sp04_Article 2 8/16/04 7:49 AM Page 752

Journal of Economic Literature Vol. XLII (September 2004) pp. 752–782

International Technology Diffusion

1 WOLFGANG KELLER

1. Introduction technology diffuses? This paper surveys what is known about the extent of international roductivity differences explain a large technology diffusion and the channels part of the variation in incomes across P through which technology is spread. countries, and technology plays a key role in International technology diffusion is determining productivity (William Easterly important because it determines the pace at and Ross Levine 2001; Robert Hall and which the world’s technology frontier may Charles Jones 1999; Edward Prescott 1998).2 expand in the future. Do countries’ incomes For most countries, foreign sources of tech- converge over time or not? The answer turns nology account for 90 percent or more of on whether technology diffusion is global or domestic productivity growth. At present, local, as several widely used models show only a handful of rich countries account for (Gene Grossman and Elhanan Helpman most of the world’s creation of new technol- 1991; Peter Howitt 2000; Luis Rivera-Batiz ogy.3 The pattern of worldwide technical and Paul Romer 1991). 4 A better under- change is thus determined in large part by standing of technical change therefore pro- international technology diffusion. What vides insights on the likelihood that certain determines the extent of technology flows less-developed countries will catch-up to between countries and the means by which rich countries. Some of the major channels for technology 1 University of Texas, Austin; NBER and CEPR. I have diffusion across countries may include inter- benefited from discussions related to this paper with Bob national trade and foreign direct investment Baldwin, Elhanan Helpman, Danny Quah, Andres Rodriguez-Claré, Dani Rodrik, Jim Tybout, and Stephen (FDI), both of which are discussed in this Yeaple. Oded Galor, Peter Howitt, Sam Kortum, Pete survey. There is evidence that imports are a Klenow, Pierre Mohnen, Jim Rauch, Carol Shiue, Dan significant channel of technology diffusion. At Trefler, and seminar participants at Industry Canada, three anonymous referees and John McMillan provided helpful the same time, the evidence for benefits asso- comments. Support by NSF grant SES 9818902 is grate- ciated with exporting is weaker. The impor- fully acknowledged. tance of FDI has long been emphasized in 2 In addition, institutions and policy have been empha- sized as well; see, e.g., Daron Acemoglu, Simon Johnson, the case-study literature, and recently that and James Robinson (2001); Abhijit Banerjee and Lakshmi Iyer (2003); and Carol Shiue and Wolfgang Keller (2004). 3 The G-7 countries (the largest seven industrialized 4 The models differ in that some predict global diffu- countries) accounted for about 84 percent of the world’s sion would lead to convergence in productivity levels while research and development (R&D) spending in 1995, for others predict convergence in productivity growth rates; example; their share in world GDP was only 64 percent. see Philippe Aghion and Peter Howitt (1998) and Charles The world’s distribution of production is much less skewed. Jones (1995), as well as the framework in section 2.

752 sp04_Article 2 8/16/04 7:49 AM Page 753

Keller: International Technology Diffusion 753

evidence has been complemented by some 1. Technology is non-rival in the sense that micro-econometric findings. the marginal costs for an additional agent Despite the global reach of computer pro- to use the technology are negligible.7 grams, there is no indication that a global pool 2. The return to technological investments of technology yet exists. This localized charac- is partly private and partly public. ter of technology suggests that an important Point 1 distinguishes technology from component of it is tacit in nature. Although rival factor inputs such as human and physi- the relative importance of international tech- cal capital; the latter can only be used by one nology diffusion appears to be increasing firm at a time, or put differently, the margin- along with higher levels of economic integra- al costs of using the same factor somewhere tion, international diffusion of technology is else are infinite. Point 2 highlights that while neither inevitable nor automatic. Domestic private returns must be strong enough to technology investments are necessary. keep innovation ongoing, technological Before discussing some of the evidence in investments often create benefits to individ- more detail (sections 6–8), section 2 pres- uals other than the inventor.8 These external ents a conceptual framework that will guide effects are called technology, or knowledge, this discussion. The following sections spillovers. As an example, the introduction review various measures of technology (sec- of one product might speed up the invention tion 3) and technology diffusion (section 4) of a competing product, because the second that are available, while section 5 discusses inventor can learn from the first by carefully the pros and cons of a number of empirical studying the product or its product design approaches that have been employed. (the “blueprint”). Section 9 summarizes the overall evidence The focus of this paper is the empirics of on international technology diffusion, and international technology diffusion and its section 10 contains a synthesis and some effect on productivity. 9 Before discussing suggestions for future work. the empirical results, I will lay out a model of international technology diffusion, based on 2. International Technology Diffusion: Jonathan Eaton and Samuel Kortum (1999) A Framework for Analysis and Kortum (1997), to help clarify the issues. Theories of endogenous technical change Consider a model with n=1,…,N coun- of the early 1990s (Aghion and Howitt 1992; tries. Output in n at time t, Ynt, is produced Grossman and Helpman 1991; Romer 1990; by combining intermediate inputs subject to and Paul Segerstrom, T.C.A. Anant, and a constant-returns-to-scale Cobb-Douglas Elias Dinopoulos 1990) emphasize two production function: 5, 6 J aspects of technology: = −1 ln()YJJnt/,∫ ln[] ZjXjdj nt ( ) nt ( ) (1)

5 See Grossman and Helpman (1995) and Aghion and where Xnt(j) is the quantity of intermediate Howitt (1998) for broader overviews. I will also discuss some related work on learning-by-doing and human-capi- tal accumulation that falls into the broad category of mod- els of knowledge accumulation. 7 Other authors have coined the terms “perfectly expan- 6 Many of these ideas have been discussed in the liter- sible” (David 1992) and “infinitely expansible” (Danny ature before; important contributors include Paul David, Quah 2001a,b) to positively define this characteristic. Giovanni Dosi, Robert Evenson, Jan Fagerberg, Richard 8 The focus of this work is on technical change driven Nelson, Keith Pavitt, Nathan Rosenberg, Luc Soete, by private, not public, research. The majority of all Sidney Winter, Larry Westphal, and others (see Fagerberg research and development (R&D) is privately funded, 1994, and Robert Evenson and Larry Westphal 1995 for although public R&D is substantial in many countries. overviews). What distinguishes the recent work is that it 9 International technology diffusion in this paper includes fully specified general equilibrium models. This refers to both technology spillovers as well as non- means that these technology effects can in principle be spillovers. As will be discussed below, the two are often estimated in a well-defined framework. difficult to separate. sp04_Article 2 8/16/04 7:49 AM Page 754

754 Journal of Economic Literature, Vol. XLII (September 2004)

input j at time t in country n and Znt(j) is the production, because by the time it has dif- quality of that input. This quality is increased fused it is dominated by a higher quality. over time through new ideas, or technolo- The stock of technologies in country n at µ = t µ· gies. The range of inputs is the same across time t is nt – ∞ nsds. The fraction of inter- countries and constant over time. Output is mediates in country n at time t that are homogeneous and tradable, while interme- below some quality thresholdq ˜ is decreasing µ diates are nontraded. in nt, because the higher is the number of Each input j is produced anywhere by a diffused technologies, the greater is the like- Cobb-Douglas combination of capital K(j) lihood that the quality in country n’s sectors φ φ and labor L(j), X(j)=K(j) L(j)1– , with is relatively high. φ ∈ [0,1]. New technologies are the result of Let Ant be the geometric mean of qual- research effort. Consider country i at time t ities across sectors. Final output (1) is maxi-

with Lit workers. If a fraction sit of them are mized when factors are evenly divided engaged in research, they create new tech- across all sectors, and it is given by β φ φ α α 1– nologies at rate itsit Lit, where it is the AntKnt [Lnt(1–snt)] . This means that Ant is research productivity, and β0 characterizes equal to output divided by a factor-share the R&D talent distribution. weighted sum of inputs, or total factor pro- Each technology has three dimensions: (i) ductivity (TFP). 11 quality, (ii) use, and (iii) diffusion time lag. To obtain a steady-state equilibrium µ First, the quality of a technology is a random where nt grows at the same constant rate in variable drawn from the cumulative distribu- all countries in the long run, we assume that = θ θ α tion F(q), F(q) 1–q , 0. This quality is the research productivity it is given by α =α =∑ N ≤ α> common to all countries to which the tech- it ( it/ t) t ; t i=1 it, 1, and 0. nology diffuses. Second, the technology can The higher is , the greater is the strength of be used only in one intermediate sector, the international research spillovers.12 which is determined randomly. Finally, steady-state relative TFP levels can Third, technologies become productive be computed as only once they have diffused. Diffusion is a 1/θ A   stochastic process, with a mean diffusion lag nt =  nt  ,,nN=−11..., , (3) –1   between country i and country n of ni , or, ANt Nt measures the speed of technology diffu- ni and world TFP growth is proportional to sion between the countries. The rate at which technologies diffuse to country n at ˙ N ε g ==n ∑ ni i sL, time t is given by ε + ii n Jgi=1 ni n = − N t − nN1,..., , (4) ˙ = JesLds1 ∑ ∫ –()ni ts , (2) nti=1 { ni 12−∞44443is444 is is4 A B where ( i/ n) and the share of labor in

research, siLi, are constant in steady-state. where A in equation (2) is the bilateral speed Equation (4) says that world growth is pro- of diffusion, and B is country i’s cumulative portional to a weighted sum of research 10 output of technologies up to time t. efforts in all countries. Even though every technology will even- 11 tually be in every country (assuming ni 0), Eaton and Kortum (1999) show that due to imperfect a given technology may not be employed in competition, TFP is not identical to Ant, but it is propor- tional to it; also see their paper for the exact definition of Ant. 12 For simplicity, we set γ to equal one. The parameter 10 Past R&D is discounted because the technologies γ is related to the question of whether the model has might have been superseded by higher-quality, more “scale effects”; see, e.g., Jones (1995), and Aghion and recent technologies. Howitt (1998) ch. 12, for more details. sp04_Article 2 8/16/04 7:49 AM Page 755

Keller: International Technology Diffusion 755

The model has a number of implications: is addressed by the major portion of the lit- 1) Foreign R&D raises domestic TFP: An erature to date, and will be reviewed in sec- increase in foreign research leads to a tions 6–9. A number of issues are important greater inflow of technologies and higher when thinking about the strength of differ- TFP. This holds in both the short and the ent diffusion channels. I will briefly turn to long-runs (equations 2, 3 respectively).13 them now. 2) Receiving technology from abroad: A How do technologies move from one given research effort abroad has a greater country to another? One possibility is effect on domestic TFP, the faster foreign through international trade in intermediate technologies diffuse to the domestic goods (Rivera-Batiz and Romer 1991; economy ( ni in equation 2). Grossman and Helpman 1991; and Eaton 3) Global sources of technology: A country and Kortum 2002).14 Employing a foreign is important in determining the world’s intermediate good in final-output produc- rate of growth if it has (i) a relatively tion involves the implicit usage of the tech- high share of the world’s research labor nology in embodied form. There is a µ and technologies (siLi and i in equation spillover in this process of international 4), and/or (ii) a relatively high rate of technology diffusion to the extent that the technology diffusion to other countries intermediate good costs less than its oppor- (the weight ni/( ni g) in equation 4 is tunity costs—which include the R&D costs increasing in ni). of product development. If trade is an International technology diffusion in this important diffusion channel, one should framework captures the notion that product expect that international technology diffu- quality innovations of country i become sion is geographically localized, because of available in country n at rate ni. Moreover, the well-documented fact that trade falls this is a spillover, because although research with geographic distance (e.g., Edward is costly in country i (it withdraws labor from Leamer and James Levinsohn 1995). In any final output production), conditional on the case, this is a relatively weak form of tech- technology having diffused, it can be used to nology diffusion, because the technology as produce final output there without incurring such is not available domestically—only the additional costs. The diffusion of technology manufactured outcome of it is. from country i also raises the productivity of Of course, international technology diffu- research in country n due to the internation- sion is not limited to the channel of trade. al research spillover (as long as 0). An In principle, just as researchers today “stand alternative, partial-spillover interpretation on the shoulders” of researchers of the past, –1 may be that the diffusion lag ni is indica- one might expect researchers in one coun- tive of the strength of costly country-n tech- try to directly benefit from research con- nology investment to acquire the technology ducted in other countries.15 The model from country i. above captures this in that research produc- In either case, it becomes clear that our tivity it is proportional to the global stock ability to give the diffusion rates ni a struc- of technologies t), which is increasing in tural interpretation that can be empirically world R&D. This is an international R&D tested is crucial for making further progress. spillover (i.e., no domestic opportunity Which are the major channels of interna- tional technology diffusion, or, what is behind the rates ? This is the question that 14 In the model above, this is not explicitly modeled, ni but it would suggest that ni is related to intermediate inputs trade; see the analysis in Eaton and Kortum (1996). 13 The same equations indicate that domestic R&D 15 See Ricardo Caballero and Adam Jaffe (1993) for raises TFP as well. empirical results in a closed economy. sp04_Article 2 8/16/04 7:49 AM Page 756

756 Journal of Economic Literature, Vol. XLII (September 2004)

costs).16 Moreover, this seems to be a technology to their subsidiaries. David stronger form of diffusion than the interme- Teece (1977) estimates that the costs of diate goods trade discussed above: if the such within-firm transfers were on average foreign technology raises the productivity of almost 20 percent of the total project costs 17 domestic researchers, it suggests full mas- in the cases he analyzed. One reason for tery of the technology, as opposed to only that is that only the broad outlines of tech- the ability to use the technology that is nology are codified—the remainder 18 embodied in the intermediate good. remains “tacit” (Michael Polanyi 1958). Which patterns of technology diffusion In this view, knowledge is to some extent should thus be expected? The first type of tacit because the person who is actively diffusion above suggests that it should follow engaged in a problem-solving activity the pattern of intermediate goods trade. cannot necessarily define (and hence International R&D spillovers appear to be prescribe) what exactly she is doing. more difficult to pin down. They are the Technology is only partially codified result of acquiring technology that is not tied because it is impossible or at least very to any particular form. In addition to costly to fully codify it. libraries, conferences and other sources, What does this imply for the channels of technology may be stored in as little as a cou- international technology diffusion? Polanyi ple of bytes, and thus sent over the internet (1958, p. 53) argues that tacit knowledge can to practically every country in the world at be passed on only “by example from master close to zero marginal cost. At the same to apprentice.” A broader view is that non- time, this does not mean that this technolo- codified knowledge is often transferred gy diffusion in fact creates equal levels of through person-to-person demonstrations technological knowledge in all countries, for and instructions (David 1992). The most the following reasons. effective way of doing this, despite tele- First, irrespective of feasibility, it is not in phone, video, and other remote methods, is the interest of the original inventor—who face-to-face interaction. This, however, has incurred the R&D cost—to send it to means that technology diffusion implies others at no charge. On the contrary, the incurring the costs of moving teacher and inventor may decide to spend additional student into the same geographic location. resources to keep the technology secret. We can summarize the implications of this Second, even if some domestic technology for the pattern of international technology becomes available abroad, it may be possi- diffusion as follows. ble to preclude others from using it by 1) The partial codified nature of technology patenting the technology. Third, even if the means that technology diffusion will be technology is non-rival and can be moved incomplete, and technology stocks in dif- from one country to another at zero mar- ferent countries will vary. Diffusion will ginal costs, operating the technology effi- tend to be more geographically localized ciently often requires making costly the higher is the non-codified share in 19 investments in terms of complementary total technology. skills (see section 8). Another issue is that technology may in 17 For transferring machinery equipment technology, this share was 36 percent (Teece 1977, p. 248); see also fact not be transferable at essentially zero Eric von Hippel (1994). cost even if all parties involved desire this, 18 See also Kenneth Arrow (1969), David (1992), and such as multinational parents providing Evenson and Westphal (1995), and references given there. 19 This follows because the costs of moving people in space are typically increasing in geographic distance; see 16 Recall that we assume γ is equal to 1. More general- von Hippel (1994) and Maryann Feldman and Frank ly, there is an international R&D spillover as long as γ>0. Lichtenberg (1997) for empirical support. sp04_Article 2 8/16/04 7:49 AM Page 757

Keller: International Technology Diffusion 757

2) Because international economic activities A drawback of R&D as a measure of tech- (trade, FDI, etc.) lead to additional nology is that it ignores the stochastic nature contacts with foreign persons who may of the process of innovation. The current possess advanced technological knowl- flow of R&D expenditures is thus a noisy edge (exporter, importer, engineers, measure of technology improvements in that researchers), this may stimulate the diffu- period. Many authors construct R&D stocks sion of (non-codified) foreign technology. from the flows using the perpetual inventory Trade and interaction with multinationals method.22 Beyond year-to-year noise, the may thus lead not only to technology diffu- return to R&D expenditures may vary sub- sion of the limited kind (technology embod- stantially across agents and over time, which ied in intermediate goods), but it may also limits comparability. One important aspect raise the probability of international R&D of this is that the return to publicly funded spillovers.20 R&D is lower than the return to privately I now turn to how recent empirical work funded R&D (Frank Lichtenberg 1993), so has studied these issues. The following two that many studies focus on business research sections discuss the data that is available and development spending. on technology (section 3) and technology Second, a patent gives its holder a tempo- diffusion (section 4). rary legal monopoly to use an innovation in a specific market at the price of public disclo- 3. Measures of Technology sure of technical information in the patent description.23 An innovation must be suffi- Technology is an intangible that is difficult ciently important to be worthy of a patent, to measure directly. Three widely used indi- which is judged by a trained official (called rect approaches are to measure (1) inputs patent examiner). Relative to R&D, patents (R&D), (2) outputs (patents), and (3) the have the advantage that patent data has been effect of technology (higher productivity). collected for a longer time (more than 150 First, internationally comparable data on years for some countries), and also poorer R&D expenditures have been published by countries have a substantial number of the Organization of Economic Cooperation patents (see WIPO 2003). and Development (OECD) since about 1965. There are some issues with using patent According to the OECD’s definition (OECD data as well. First, a small number of 2002), only about two dozen relatively rich patents accounts for most of the value of all countries report substantial amounts of patents. This means that simple patent R&D, because the definition captures prima- counts may not measure technology output rily resources spent towards innovation, and well. Recent work has addressed this issue not those spent on imitation and technology in part by using citation-weighted patent adoption. Technology investments of middle data (see Adam Jaffe and Manuel and poor countries can therefore typically not Trajtenberg 2002).24 Second, the patent be analyzed using R&D data.21 decision is an act of choice on the part of the firm, and a large set of innovations is not 20 This process could also in part be sequential: (1) using foreign technology in embodied form, and gradually (2) learning of the technology per se, (3) imitation, and (4) 22 This requires an assumption on the rate of R&D (original) innovation. depreciation; standard values range between 0 and 10 per- 21 However, R&D data becomes more widely available cent; see Zvi Griliches (1995). as countries’ incomes are rising. There is also increasingly 23 See Griliches (1990) for more discussion. information on poorer countries because surveys encom- 24 Typically, a patent contains one or more references to pass the R&D conducted by affiliates of multinational other patents that indicate which earlier knowledge was companies located abroad; see e.g. NSF (2004) for R&D used to come up with the technology underlying the pres- expenditures of U.S. firms in China. The main OECD ent patent application; these references are called patent R&D statistics are on a geographic, not ownership, basis. citations. sp04_Article 2 8/16/04 7:49 AM Page 758

758 Journal of Economic Literature, Vol. XLII (September 2004)

ever patented.25 And third, if technology is in Because of these difficulties in computing part non-codifiable, as argued above, patent TFP, researchers have pursued a number of statistics will necessarily miss that part. strategies. One is to consider changes in TFP The third measure of technology dis- as opposed to TFP levels. This will help in cussed here is total factor productivity identifying technology (or rather, technical (TFP). The idea, well-known since the change) if spurious factors do not change over 1950s, is that if one subtracts from output time, or more generally, if they change less the contribution of inputs such as labor and than technology. For example, in Katayama, capital, the remainder is due to the factor Lu, and Tybout’s (2003) case from above, if a “technology.”26 A simple example is the term firm faces higher adjustment costs to chang- A in the Cobb-Douglas production function ing its mark-up, this will reduce the mark-up YAK L1– . Other TFP measures are more variability in equilibrium. A second strategy general and have certain desirable proper- has been to employ TFP measures in studies ties that are important for comparability of technical change together with data on (e.g. the “superlative” index proposed by R&D (e.g., Griliches 1984). By establishing a Richard Caves, Laurits Christensen, and relationship between TFP changes and its Erwin Diewert 1982). presumed major cause, R&D spending, the In contrast to R&D and patents, TFP is a likelihood of measuring changes in technology derived measure of technology, as it is com- inappropriately is substantially reduced. puted from data on inputs and output. This I now turn to measures of technology 28 introduces measurement error and perhaps diffusion and spillovers. biases, because the appropriate data on 4. Measures of International inputs and outputs is rarely, if ever, available. Technology Diffusion Hajime Katayama, Shihua Lu, and James Tybout (2003), for example, show that the The diffusion of technology involves both use of (1) real sales revenues, (2) depreciat- market transactions and externalities (see ed capital spending, and (3) real input above), and obtaining data on the former is expenditures; instead of (unavailable) data fairly straightforward. For instance, firms on the physical quantities (1) of output, (2) of make royalty payments for their use of capital, and (3) of intermediate inputs, as is patents, licenses, and copyrights, and this frequently done, will often confound higher data is available for major countries in the productivity with higher mark-ups. 27 Other international services balance (e.g., OECD factors might thus contaminate the use of 2003). Many economists believe, though, TFP as a measure of technological efficiency, that most international technology diffusion which ultimately goes back to the concern occurs not through market transactions but that TFP is constructed as a residual, and instead through externalities (spillovers).29 may potentially capture a host of spurious influences (Moses Abramovitz 1956). 28 Sometimes firm-level datasets provide information on the fraction of computer-controlled machinery, or whether a firm satisfies certain technological standard; either of these would also provide information about the 25 The fact that there is variation in the propensity to technological sophistication of the firm. patent in several dimensions—for instance, it is profitable 29 One reason is that technology is not fully codifiable, to patent only in a large-enough market—could also pro- so that complete contracts cannot be written. Another rea- vide important information. son is market failure in the market for technology due to 26 See Charles Hulten (2000) for more details on TFP. asymmetric information: the buyer does not know (the 27 This could have major social welfare implications as productivity of) the technology, while the seller cannot consumers benefit from higher productivity but not from commit to truthful claims about it. This is considered to be higher mark-ups. See also Tor Klette and Griliches (1996), a major reason why international technology transfers are as well as Eric Bartelsman and Mark Doms (2000) for a often internalized within firms (between multinational recent analysis of micro productivity data. parent and subsidiary); see, e.g., William Ethier (1986). sp04_Article 2 8/16/04 7:49 AM Page 759

Keller: International Technology Diffusion 759

Naturally, data on spillovers does not exist. us correlations but not causal effects. These, Measures that are related to it do exist, but as well as issues in other approaches, such as typically they capture spillovers only partial- estimates of spillovers as the parameter ni in ly, because the measures do not account for the model of section 2, are discussed in the costs of acquisition (learning). For instance, following section. if one patent application cites an earlier patent, this generally indicates that the 5. Empirical Methods applicant has benefited from the earlier 5.1 Case Studies patent.30 At the same time, it is impossible to know how large these benefits are net of the Case studies can offer a rich description of learning costs that the patent applicant had the setting and the major factors that deter- to incur. mine international technology diffusion. For Among the different methods that try to instance, the paper by Felipe Larrain, Luis measure international spillovers, the largest Lopez-Calva, and Andres Rodriguez-Claré set of papers employs international R&D (2000) analyzes whether Intel’s foreign spillover regressions. In one set of papers, if direct investment into Costa Rica in the R&D of firm j is positively correlated with 1990s generated any technology spillovers. TFP in firm i, all else equal, this is consistent The authors have spoken to the main actors, with international technology spillovers from including Intel’s top management and the firm j to firm i (e.g., Keller 2002a).31 A vari- government of Costa Rica, about their con- ant of this approach replaces TFP by the cerns and motivations. This is informative. number of patents (e.g. Lee Branstetter The major limitation of case studies is that 2001b), and Giovanni Peri (2002) presents a one does not know how general one particu- hybrid approach by relating patents in region lar case is. Nevertheless, there seems to be a i to patents in other regions, where the latter role for case study research, especially if the is instrumented by R&D expenditures. study is—as in the Intel case—backed up by There are two alternatives to this basic quantitative evidence.32 approach, a generalization and a simplifica- 5.2 Econometric Studies tion. In the former, a particular channel of technology diffusion is added to the analysis. 5.2.1 Association Studies David Coe and Elhanan Helpman (1995), In association studies, authors ask e.g., analyze the relationship between pro- whether a specific foreign activity (FA) ductivity and foreign R&D conditional on leads to a particular domestic technology imports from that foreign country. The other outcome (DTO): alternative to the international R&D spillover regressions is to relate productivity DTOƒ(X,FA)u (5) not to foreign R&D, but to other measures In equation (5), X is a vector of control vari- of foreign activity; Brian Aitken and Ann ables, and u is a regression error. For Harrison (1999), e.g., study the correlation of instance, Aitken and Harrison (1999) exam- inward FDI and domestic firm productivity ine the importance of foreign direct invest- (so-called FDI spillover regressions). ment (FDI) as a source of international A common concern in these studies is that, technology spillovers; here, FA is the indus- for various reasons, the estimate might give try share of employment in foreign-owned firms, and DTO is increases of domestic firm 30 It is not a perfect indicator because sometimes cita- productivity. tions are added by the patent examiner, not the applicant. 31 This approach goes back to the closed-economy work by Griliches (1995) and Frederic Scherer (1984). 32 See section 6.2 for more on this particular case study. sp04_Article 2 8/16/04 7:49 AM Page 760

760 Journal of Economic Literature, Vol. XLII (September 2004)

This approach is based on economic theo- by including time fixed effects, as long as it is ry in the following sense. Often, there are the case that the business cycle is common to several models that have been proposed to the entire sample. More difficult is the ques- explain, in this case, FDI spillovers, and tion whether trend variables are appropriate, model-specific evidence often does not yet given the time-series properties of the data, in exist. Association studies try to shed light on particular whether it is stationary. This is an the most interesting models by proposing important issue because including trend vari- what might be the common reduced-form ables can have a major impact on the results equation of all these FDI spillover models. in this area. 35 Panel unit root tests and coin- In order to accommodate several models, tegration analysis can be applied, but that the framework cannot be very specific. This approach is still relatively new, and results explains why authors of FDI spillover stud- vary.36 This suggests trying out alternative ies employ a simple measure such as the for- approaches; if results change dramatically eign employment share. If FDI has a major depending on the assumption on the data effect on domestic firm productivity, it will generation process, this warrants at least be picked up by foreign employment no some discussion. matter what the particular mechanism is. Second, a standard approach in the pres- The generality is attractive, but it also pre- ence of unobserved heterogeneity is to use cludes a precise interpretation of the results. fixed effects.37 The consequence of course is Moreover, by using a foreign activity (FA) to reduce the variation in the dependent instead of a foreign technology variable, this variable that is left to explain. As such, fixed approach is prone to estimating technology effects tend to be of little interest, as they diffusion only with some error.33 This mat- rarely capture the economics that is ters because calculating a proper causal involved. Moreover, the differencing that is effect (instead of a correlation) becomes implied by the fixed effects can exacerbate more difficult as the possible causes of spu- measurement error problems (Zvi Griliches rious correlation multiply. This could be due and Jerry Hausman 1986). These considera- to business cycle effects or to unobserved tions suggest keeping the number of fixed heterogeneity. If productivity were exoge- effects low. At the same time, there is often nous and FDI would vary inversely with some uncertainty about the ‘true’ model, trade costs across industries, a positive coef- and if a low number of fixed effects means ficient of FDI on productivity would suggest omitted variables bias, the results can be spurious FDI spillovers if high-trade cost seriously misleading. In general, it will be industries would exhibit on average higher worth trying alternative specifications. productivity, for example.34 Endogeneity is Fixed effects will of course not work as a another problem; here, it could take the control for unobserved heterogeneity if the form of foreign investors choosing to move FDI into industries that have high rates of 35 For instance, Coe and Helpman (1995) report that productivity growth (those industries are including a generalized time trend in their regression typically also relatively profitable). reduces the international spillover coefficient by about 60 percent, and using conventional standard errors, the How can these issues be addressed? First, parameter might be insignificantly different from zero spurious business-cycle effects can be avoided (table B.1 and table 3, coefficient on m∗Sf). 36 For instance, critical values are dependent on the exact specification of the intercept and trend components, 33 Here, using the foreign employment share as the as well as how much cross-sectional heterogeneity is measure of FDI may be justified because multinational allowed for; see G.S. Maddala and In-Moo Kim 1998) for enterprises tend to be R&D-intensive and share their an introduction, and Chris Edmonds (2000) for an appli- technology between parent and subsidiaries. cation to spillover regressions. 34 Lael Brainard (1997) finds that FDI is relatively high 37 Time-differencing is often used for the same pur- in high-trade cost industries. pose; see e.g. Keller (2000). sp04_Article 2 8/16/04 7:49 AM Page 761

Keller: International Technology Diffusion 761

heterogeneous effects are actually not time- estimated to begin with, and moreover, most invariant. Not much is known on the extent variables that are uncorrelated with the to which this has led to biases in the existing error term tend to be not particularly corre- studies. At any rate, the dynamic industry lated with R&D, especially at the industry- framework of Steven Olley and Ariel Pakes or micro-level. (1996) does not rely on this assumption—it The lack-of-instruments problem should allows for time-varying heterogeneity—and become smaller because some recent IV it has been employed in a number of micro techniques use primarily lags of right hand studies.38 side variables as instruments. The instru- The third point, endogeneity, has been ments are valid as long as there is not too recognized in the literature, but it is rarely much persistence—which can be tested fully addressed. For instance, Keller for—and the robustness of the approach (2002a) splits his sample into two, the high- exceeds earlier techniques because it uses and low-R&D expenditure sectors, based both the variables’ levels and their on the fact that four out of his twelve indus- changes (see Richard Blundell and tries account for more than 80 percent of all Stephen Bond 1999). R&D. For a variety of reasons, endogeneity Two papers that have used IV estimation problems are much more likely to arise in recently are Rachel Griffith, Stephen the former than in the latter sectors. Redding, and Helen Simpson (2003) and Keller’s result for the low-R&D sample Keller and Yeaple (2003). Both of these turns out to be similar to that for the whole papers study the importance of technology sample. In general, association studies con- spillovers associated with FDI.39 The duct often a wide number of robustness Griffith et al. (2003) paper studies the influ- analyses. These analyses are hardly defini- ence of foreign-owned subsidiaries in the tive in dealing with endogeneity: they are United Kingdom. Their instruments, the sequential in nature (looking at one sus- economic conditions in France and the pected problem at a time), they proceed United States, are quite plausible, because under assumptions that might not hold these are the countries from which many (e.g., lagging a regressor while assuming of the U.K. subsidiaries come. Using over- that there is no serial correlation), and for identification tests, the authors find that other reasons. Nevertheless, such robust- endogeneity is not an issue in their sample. ness analysis can be very helpful in reveal- This may be somewhat surprising at first, ing whether endogeneity might have a perhaps suggesting that the instruments are first-order effect in a particular context. not as good as one might hope for. However, Endogeneity issues can in principle be Keller and Yeaple’s (2003) IV analysis does fully addressed by using instrumental vari- not suggest an endogeneity bias either. able (IV) techniques. So far these tech- These authors use industry level transport niques have not been used widely in this costs and tariffs as instruments for FDI literature, although that seems to be chang- industry variation in the United States. ing now. In general, the application of IV Keller and Yeaple’s IV estimates suggest a techniques has been limited because finding higher effect through FDI spillovers than is good instruments for technology variables obtained with OLS. This is the opposite of such as R&D stocks is difficult. These are the suspected endogeneity bias, and could be the result from a FDI variable that does not 38 See, for instance, Sourafel Girma and Katharina measure too well the source of technology Wakelin (2001), Garrick Blalock and Paul Gertler (2002), spillovers. and Wolfgang Keller and Stephen Yeaple (2003). On Olley and Pakes’ (1996) method, see also Zvi Griliches and Jacques Mairesse (1998). 39 See section 8 for a broader discussion of this. sp04_Article 2 8/16/04 7:49 AM Page 762

762 Journal of Economic Literature, Vol. XLII (September 2004)

Despite these results, it is too early to distinct from n, the range of intermediate argue that endogeneity does not play a major goods produced in the country. The latter is role in many studies. More research is clear- augmented through R&D (denoted ); if ly needed. Other approaches that should be intermediate goods do not become obsolete, used more broadly are difference-in-differ- the range of intermediate goods at time T is ence and other control-group estimations, given by the cumulative resources devoted T τ τ especially when a clear regime shift (“natural to R&D, nT ST 0 ( )d . The goods x(s) experiment”) can be identified. are best thought of as differentiated capital I now turn to a second class of empirical goods, produced with foregone consump- papers that has been employed in the econo- tion. If the x(s) are symmetric and linearly metric literature. produced with foregone consumption, one can write the stock of capital as 5.2.2 Structure Studies n k 0 x(s)ds nx, which can be used to obtain These studies incorporate more of the a reduced-form expression for output as structure of the underlying model than asso- zA(ne) l k1– . (9) ciation studies. I distinguish between two ≡ z types of structure studies. Generally, the first Defining TFP as f 1− , one obtains set is given by lk ln flnAlnne. (10) DTOƒ(X,M,FT)u (6) Equation (10) shows that TFP is positively where the foreign technology variable, FT, related to the range of intermediate prod- replaces the foreign activity variable in ucts employed in this country (Ethier 1982). equation (5), and the specification adds a In a symmetric model with C countries specific channel, or mechanism of diffusion and no trade barriers, in equilibrium all (denoted by M). countries will employ all intermediates from An influential example is the study by Coe all countries. Given that the intermediate and Helpman (1995). These authors test the designs are produced through national prediction of the trade and growth models of R&D, the range ne will be the same in all Grossman and Helpman (1991), and Rivera- countries, equal to the world’s cumulative Batiz and Romer (1991), in which foreign R&D spending: R&D creates new intermediate inputs and C C perhaps spillovers that the home country can e == nnSt ∑ ct∑ ct . (11) access through imports. Assume that output c=1 c=1 is produced according to Coe and Helpman’s (1995) test of the zAl d1– , 01, (7) model recognizes that there is both country where A is a constant, l are labor services, heterogeneity as well as trade barriers by and d is a CES aggregator of differentiated separating domestic and foreign R&D: d f f intermediate inputs x of variety s ln fc c lnSc lnSc c, (12) 1  ne  1− − where country c’s so-called foreign knowl- = 1 f dxsds ∫ () , (8) edge stock S is defined as the bilateral   c 0 import-share weighted R&D stocks of its where ne is the range of intermediate goods trade partners: 40 that are employed in this country. It can be f = SmSc ∑ cc'' c , (13) cc'≠ 40 This follows Romer (1990) and Grossman and Helpman (1991); see also Keller (2000); time subscripts This captures the prediction of Grossman are omitted when no ambiguity arises. and Helpman’s (1991) model that if a country sp04_Article 2 8/16/04 7:49 AM Page 763

Keller: International Technology Diffusion 763

imports primarily from high-R&D partner  J ≤++x e c countries, it is likely to receive relatively 10if Xit eit∑ j ln(AVCit− j ) much technology embodied in intermediate  j=1  J goods, which should be reflected in a higher = +−0 j + yit  ∑()FFyit− j it productivity level; and vice versa.  j=1 This approach has been influential 0 otherwise, because of its plausibility, simplicity, and ver-  (14) satility, and has been widely used to examine  alternative channels of international diffu- and an autoregressive cost function, sion; for instance, Frank Lichtenberg and J =+∑ k + e Bruno van Pottelsberghe de la Potterie ln(AVCit )0 j ln( Kit− j ) ln( et ) j=1 (2000) have examined FDI by substituting J J c y bilateral measures of FDI instead of +++∑ln(AVC−− ) ∑ y v (15) 41 j it j j it j it imports. The approach is more specific j=1 j=1 about the mechanism, which helps in inter- preting the results. At the same time, such where yit is the export indicator of plant i in regressions estimate different models only period t, Xit is a vector of exogenous plant characteristics, et is the exchange rate, AVCit partially; what determines R&D, for 0 are average costs, Kit is capital, and F and instance, is typically not modeled, which j means that endogeneity could be an issue. F are sunk costs terms (of export market Given the partial equilibrium nature of the participation). estimation, authors should and generally do Equation (14) states that one only sees a consider alternative specifications. plant exporting if the profits from doing so The second set of studies in this section are greater than from not exporting (the puts relatively less emphasis on a particular latent threshold is expressed in terms of model of growth and diffusion, and more observables). Equation (15) estimates emphasis on econometric specification and whether past exporting experience reduces estimation. A good example is the paper by current cost (captured by the parameters y Sofronis Clerides, Saul Lach, and James j ), conditional on past costs and size (prox- Tybout (1998), where the authors provide ied by capital). These learning-by-exporting evidence on learning externalities from parameters are not constrained in a signifi- exporting using micro data from Columbia, cant way—recall that yit–j are 0/1 indicator Morocco, and Mexico. Clerides, Lach, and variables—which is probably a good starting Tybout (1998) extend the sunk cost model of point given that it is unknown how learning exporting due to Richard Baldwin (1989), from exporting works (if indeed it exists). Avinash Dixit (1989), and Paul Krugman Clerides, Lach, and Tybout (1998) employ alternative estimation methods in their (1989) to include the possibility that export- 42 ing experience lowers the costs of produc- analysis. Moreover, the econometric evi- tion. Clerides et al. take account of the dence is combined with descriptive evidence important point that it is on average the in form of over-time plots of average costs already-productive firms that self-select into for different types of plants (exporters, non- the export market. They do this by estimat- exporters, etc). The emphasis throughout is ing simultaneously a dynamic discrete to distill the results that emerge consistently choice equation that determines export from all empirical approaches (they are dis- market participation cussed in section 6.1.2 below). I would

42 This includes full information maximum likelihood 41 The approach of computing a weighted sum of R&D and generalized methods of moments, the latter with a goes back to Griliches (1979). generalized cost function. sp04_Article 2 8/16/04 7:49 AM Page 764

764 Journal of Economic Literature, Vol. XLII (September 2004)

expect more work along these lines to be diffusion lag to other countries is exponen- very fruitful. tial. Within the context of a given model, such assumptions are very difficult to test. 5.2.3 Empirical Analysis with General It means that this empirical work does not Equilibrium Models so much test or select among several mod- Eaton and Kortum (1996, 1997, 1999) els, as it estimates one particular model. have studied international technology diffu- Second, the models are usually too compli- sion using general equilibrium models. In cated for estimation to identify all model their models productivity growth is related parameters. If, as is common practice, a to increases in the quality of intermediate number of parameters are fixed based on goods, which is key to Aghion and Howitt’s values from earlier studies, this means that (1992) quality ladder model of growth. the results are partly simulated, and not Eaton and Kortum add a process that gov- estimated.44 erns technology diffusion between coun- Third, it is often difficult to judge how tries (see section 2 above). They then empirically successful a model is. One meas- explore the quantitative relationships ure of success is the difference between between R&D, technology diffusion, and actual versus predicted values for the domestic productivity. endogenous variables. If a model predicts Eaton and Kortum’s work is important productivity levels that lie within an error because instead of focusing on the reduced- margin of 15 percent of the actual levels, for form relationship between a subset of vari- instance, how much of this is due to “free” ables, they use their full model for making model parameters (i.e., those that are set ex- predictions on all (endogenous) variables: ante)? With those parameters, how much endogeneity as in the single-equation studies worse would a different, perhaps simpler above is therefore no longer an issue. Eaton model fare? And what if the set parameters and Kortum study both the long-run equilib- are adjusted to maximize the fit in the alter- rium and the models’ predictions for the native model? Peri (2002), for instance, uses transitional dynamics (Eaton and Kortum a partial equilibrium approach to estimate 1999 and 1997, respectively). Eaton and technology diffusion in a model similar to Kortum (1996) shed additional light on the Eaton and Kortum (1996) who work in a rates of diffusion in their model by using general equilibrium framework. With N reduced-form techniques. Several of their countries, NN equations pin down bilater- papers study comparative-static changes in al technology diffusion, and the implied the long-run equilibrium, which is especially technology stocks determine (N1) relative useful for economic policy analysis (Eaton, productivity levels, as in equation (3) above. Eva Gutierrez, and Kortum 1998).43 According to Peri (2002), the technology There are costs to this approach as well stocks do not have a significant effect on pro- though. First, in order to arrive at a frame- ductivity. In Eaton and Kortum (1996), that work that can be analyzed, typically some relationship is one element of their model’s strong assumptions have to be made. In general equilibrium structure, which means Eaton and Kortum (1999), for instance, the that it could be that the structural assump- quality of technology that is discovered in a tions have too strong an influence on the country is a random variable with a Pareto empirical results. distribution, while the distribution of the

44 In addition, it can be difficult to find other studies that have estimated a particular model parameter, for 43 See also Eaton and Kortum (2002); Andrew Bernard instance θ, which governs technology discovery in Eaton et al. (2003); and the discussion in 6.2 below. and Kortum (1999). sp04_Article 2 8/16/04 7:49 AM Page 765

Keller: International Technology Diffusion 765

It is important to ask which part of the section 2—with the Ricardian model of variation in the data, or which assumptions trade due to Rüdiger Dornbusch, Stanley in the model, are primarily responsible for Fischer, and Paul Samuelson (1977).46 In the results. Despite the sensitivity analysis Eaton and Kortum’s model, trade augments that may be presented, the informed reader a country’s production possibilities for the will typically find this question impossible to classic Ricardian reason: trade gives access answer, and this reduces the usefulness of to foreign goods or, implicitly, technologies. the general equilibrium results.45 By specializing in their respective compara- To summarize, the different empirical tive advantage goods, countries can gain approaches each have their advantages and from trade in the sense that given a country’s disadvantages. There is a case for adopting a resources, the efficient level of output with broad estimation framework, as in associa- trade is higher than without trade. At the tion studies, if there is model uncertainty, same time, there are no spillovers in this and this case is stronger if endogeneity model, in the sense that importers pay the issues can be addressed. It is useful to competitive price and importing has no impose more structure as more information effect on innovation.47 becomes available, because the structure Eaton and Kortum assume that unit trans- aides interpretation. General equilibrium port costs are increasing in geographic dis- models allow us to conduct counterfactuals, tance. This implies that the price of which is crucial for policy and welfare analy- intermediate (or equipment) goods in sis, and they will be most influential if the remote countries is relatively high, or, equiv- sources of variation that underlie the results alently, that productivity in these countries is are clear. relatively low. These effects are shown to be The following section discusses the quantitatively important, as differences in importance of different potential channels the relative price of equipment account for of international technology diffusion. 25 percent of the cross-country productivity differences in a sample of 34 countries (Eaton and Kortum 2001). 6. Channels of International However, it is not clear at this point Technology Diffusion whether this provides strong support for 6.1 Trade imports as a major channel for technology diffusion. This is because the equipment 6.1.1 Imports and Technology Diffusion goods prices predicted by Eaton and First, I turn to the evidence on diffusion Kortum’s (2001) model are inversely related through intermediate input imports. In to those reported in Robert Summers and Eaton and Kortum (2002, 2001), the authors Alan Heston’s International Comparison have combined the structure of technology Program of Prices data (CIC 2003): while diffusion in Eaton and Kortum (1999)—see according to Summers and Heston’s data, rich countries have higher equipment prices than poorer countries, Eaton and Kortum’s 45 In comparison, it is straightforward to assess partial- equilibrium work. For instance, according to Coe and model predicts the opposite (Eaton and Helpman (1995), bilateral import patterns are important Kortum 2001, figure 7). for explaining variation in productivity; see equation (13). A simple test of that is to drop the import shares in (13), and see whether it leads to a major reduction in explained 46 See also the extension by Bernard et al. (2003) to variation of productivity (Keller 1998; see section 6.1). plant-level analysis. Empirical work on technology diffusion in general equilib- 47 As far as I know, the literature to date lacks a model rium has not been subjected to the same level of scrutiny, in which imports have a direct effect on technology dif- presumably because the nature of the framework has more fusion, at least for the multi-country case with sectoral to do with simulation, and less with testing to begin with. differences. sp04_Article 2 8/16/04 7:49 AM Page 766

766 Journal of Economic Literature, Vol. XLII (September 2004) Summers and Heston’s price data might denoted by c'c, in place of the actual bilater- be imperfect, but it is unlikely that revisions al import shares to create the counterfactual f∑ of this data will reverse how equipment foreign knowledge stock Sct c'≠c cc'Sc't. prices correlate with income. At this point, Using this alternative foreign R&D variable Eaton and Kortum’s model, although yields similarly high coefficients and levels of appealing because it suggests plausibly large explained variation as the regressions using f 48 productivity effects come from importing Sct (that is, imports data). Given that foreign technologies, seems to require a import shares are not essential for Coe and major modification in order to be consistent Helpman’s (1995) results, their analysis does with international equipment goods price not allow us to draw strong conclusions data. The latter is not the only core predic- regarding the importance of imports as a tion of the model, but it appears to be vehicle for diffusion.49 important enough to put at least a question A number of authors have made progress mark behind the estimated quantitative by examining the international R&D importance of trade for technology diffusion. spillover regressions further. Bin Xu and Second, there is evidence on the impor- Jianmao Wang (1999) emphasize that tech- tance of imports that comes from interna- nology diffusion in recent trade and growth tional R&D spillover regressions. Based on models is associated specifically with differ- the framework laid out in section 5.2.2 entiated capital goods trade. This is in con- above, Coe and Helpman (1995) relate TFP trast to the trade data Coe and Helpman to domestic and foreign R&D, (1995) use to construct their import shares d f f (from overall trade). Xu and Wang (1999) lnfSS=+ ln + ln +ε , (12’) ct c ct ct ct show that this distinction matters: the capital f where Sct is defined as the bilateral import- goods-foreign R&D variable accounts for share weighted R&D stocks of its trade part- about 10 percent more of the variation in f∑ ners, Sct ≠ mcc'Sc't. A positive effect from productivity than does Coe and Helpman’s c' c f the foreign R&D variable Sct would imply analysis, and it also performs better than that a country’s productivity is increasing in Keller’s (1998) counterfactual variable. 50 the extent to which it imports from high- as It has also been noted that the foreign f opposed to low-R&D countries, supporting R&D variable Sct captures only current- the hypothesis that imports are a channel of period bilateral trade; it is clearly possible technology diffusion along the lines of the though that country A benefits from country models discussed in Grossman and C’s technology without importing from this Helpman (1991). In a sample with 22 source, if country C exports to country B, OECD countries, Coe and Helpman (1995) which in turn exports to A. Olivier estimate a positive and quantitatively large Lumenga-Neso, Marcelo Olarreaga, and effect from import-weighted foreign R&D. Maurice Schiff (2001) use a specification Similar effects are found for technology dif- fusion from highly industrialized to 77 less 48 Alternatively, Keller (1998) sets all µcc' equal to 1, developed countries (Coe, Helpman, and which produces similar results. This confirms that the import shares are inessential for the results, whether or not Alexander Hoffmaister 1997). they are truly random (see Keller 1997, 2000; and Coe and There are some reasons to remain skepti- Hoffmaister 1999). cal here as well. First, the analysis of Keller 49 Keller’s (1998) analysis has the same implication for Coe, Helpman, and Hoffmaister’s (1997) paper. (1998) has shown that the import shares in 50 Keller (2000) uses specialized machinery imports the construction of the foreign R&D variable data at the industry level in a related analysis. His results S f are not, in fact, essential to obtain Coe show that technology diffusion through imports does affect ct productivity growth for high skewed trade patterns (e.g. and Helpman’s (1995) results. Specifically, for Canada, which imports to about 70 percent from the Keller (1998) uses randomly created shares, United States), suggesting that the effect is nonlinear. sp04_Article 2 8/16/04 7:49 AM Page 767

Keller: International Technology Diffusion 767

that captures such indirect R&D spillovers, There is abundant evidence that in a given and show that it performs better than Coe cross-section, exporters are on average more and Helpman’s (1995) and Keller’s (1998) productive than nonexporters (e.g., Bernard models. These results are consistent with and Jensen 1999; Clerides, Lach, and Tybout the importance of dynamic effects from 1998; Mary Hallward-Driemeier, Giuseppe imports, but more research in an explicitly Iarossi, and Kenneth Sokoloff 2002).51 That dynamic framework is needed to learn more does not settle the issue of causality, howev- about this. er: are exporting firms more productive The importance of imports for technology because of learning effects associated with diffusion has also been assessed with patent exporting, or is it rather the case that firms citation data. Frederic Sjöholm (1996) stud- that are more productive to begin with start ies citations in patent applications of exporting? The conventional wisdom today Swedish firms to patents owned by foreign is that learning-by-exporting effects are non- inventors. Controlling for a number of other existent. This position is consistent with cur- correlates and also conducting an extreme- rent evidence, but at the same time there are bounds analysis, Sjöholm finds a positive a number of issues which in my view are correlation between Swedish patent cita- worth further research before the question tions and bilateral imports. This is consistent should be considered settled. with imports contributing to international An important paper is by Clerides, Lach, knowledge spillovers. and Tybout (1998). These authors study Overall, the evidence points to a significant manufacturing plants in Columbia, role for important in international technolo- Morocco, and Mexico during the 1980s to gy diffusion. However, the various strands of find evidence on learning-by-exporting the literature leave still some questions open, effects. Their framework is based on a model and we do not have yet a firm estimate of with sunk costs of (export) market entry, the quantitative importance of imports for leading to a dynamic discrete choice equa- international technology diffusion. tion for participation and an autoregressive cost function as a performance measure (see section 5 above). 6.1.2 Exports and Technology Diffusion: Clerides, Lach, and Tybout show results Is There Learning-By-Exporting? for each country separately, and also by A major question is whether firms learn major industry. In general, they tend to about foreign technology through exporting show no significant effects from past export- experience. There is anecdotal evidence ing experience on current performance y claiming that firms do benefit from interact- (this is parameter j in equation 15). ing with foreign customer, for instance Clearly, this does not lend support for the because the latter impose higher product existence of strong positive learning-by- quality standards than domestic customer, exporting effects. In fact, to the extent that while at the same time providing informa- Clerides et al.’s estimates are significant, tion on how to meet the higher standards. they go into the wrong direction (exporting Case studies of the export success of a num- raising costs). It would be surprising if ber of East Asian countries starting in the indeed there would be negative learning 1960s are particularly strong in their effects, and the authors give a number of emphasis on learning-by-exporting effects (Yung-Whee Rhee, Bruce Ross-Larson, and 51 Garry Pursell 1984). The question is Note that the evidence on learning-by-exporting effects discussed in the following comes from micro stud- whether this evidence can be supported ies, while the evidence on diffusion associated with imports with econometric evidence. discussed above is based on more aggregate studies. sp04_Article 2 8/16/04 7:49 AM Page 768

768 Journal of Economic Literature, Vol. XLII (September 2004)

plausible reasons of why this finding may exporters.54 It is plausible that this 10 per- have to be discounted. Another interpreta- cent survival probability difference is indica- tion of the generally insignificant estimates tive of higher productivity growth for may be that the estimation framework is exporters than nonexporters, because plants demanding too much of the data. However, tend to fail because their productivity Clerides et al.’s plots of average cost before growth is low. This suggests that the overall and after export market entry seem to sup- difference in productivity growth between port their main estimation result of no exporters and nonexporters may be larger evidence for learning-by-exporting effects. than 0.8 percent, although at this point it is While learning-by-exporting has been not clear how much larger. emphasized to be important primarily for Some of these estimates come from a rel- low- and middle-income countries’ firms, atively small number of years, and there there is in principle no reason why it is lim- may be an argument that this time horizon ited to these countries, especially given the is too short to see major learning-by-export- firms’ heterogeneity in terms of productiv- ing effects. Alternatively, Hallward- ity in any given country. Bernard and Driemeier, Iarossi, and Sokoloff (2002) Jensen (1999) study the learning-by- focus on the time before entering the export exporting question using data on U.S. market. These authors use data from five firms. This has the advantage that the sam- Southeast Asian countries to show that firms ple is relatively large and there is compara- which eventually become exporters make tively much experience with data collection more investments to raise productivity and and preparation, which may result in lower the quality of their goods than firms that measurement error. plan to stay out of the export market. This is Unlike Clerides, Lach, and Tybout (1998), plausible, but if these investments Bernard and Jensen (1999) do not model require—which is likely—real resources, export market participation explicitly. those need to be subtracted from any learn- Instead, they study the performance of the ing effects the firms receive after they have 52 different sets of firms separately. Bernard entered the export market.55 Moreover, and Jensen estimate that labor productivity given that the productivity increases pre- growth is about 0.8 percent higher among date the firm’s entry into the export market, 53 exporters than nonexporters. This esti- at best these are indirect learning-by- mate is fairly small, and it becomes even exporting effects. smaller (and insignificant) for longer time The analysis has shown that there is no horizons. econometric evidence for a strong learning- The estimate of 0.8 percent, however, from-exporting effect. At the same time, appears to be a downward-biased estimate there are a number of issues that still need to of the learning-by-exporting effect because be addressed. First, it is puzzling that the it comes from an analysis conditional on plant survival. In fact, Bernard and Jensen show that conditional on size, exporters are 54 Size is the main predictor of survival in recent indus- try-equilibrium models (e.g. Olley and Pakes 1996), 10 percent more likely to survive than non- because a small firm might have to exit after only one bad shock (or, a small number of successive bad shocks), 52 Four types of firms can be distinguished: exporters, whereas a large firm has substance enough to weather a nonexporters, starters (plants that start exporting), and longer succession of bad shocks. quitters (plants that stop exporting). 55 This point is related to the fact that none of the esti- 53 Bernard and Jensen’s estimates using TFP instead of mates I discuss are claimed to be spillover effects; rather, labor productivity are lower, but the labor productivity fig- they are “learning” effects, which could be costly to ures are preferred in this case. The TFP measure is a sim- acquire. Clerides, Lach, and Tybout (1998) do estimate ple regression residual that is fallible to a number of spillovers, those that might accrue to other plants; this problems (e.g. Griliches and Mairesse 1998). evidence is mixed. sp04_Article 2 8/16/04 7:49 AM Page 769

Keller: International Technology Diffusion 769

econometric evidence is so strongly at odds (see Jonathan Haskel, Sonia Pereira, and with the case-study evidence. Second, there Matthew Slaughter 2001). could be heterogeneity across industries What is the evidence on FDI spillovers? A masking strong but industry-specific learning number of surveys have recently concluded effects, especially with respect to high-tech- that there is no evidence for substantial FDI nology products.56 Third, the analysis could spillovers (Gordon Hanson 2001; Holger be improved if we knew more on both the Görg and David Greenaway 2002).58 export destination and the exporter, instead However, this view, based primarily on a of simply an indicator variable (exporting number of micro-level productivity studies, yes/no). For instance, to which firms, in seems to be overly pessimistic. First of all, which countries, do the exports go? What are some more recent evidence for the micro the characteristics of these foreign firms— productivity approach suggests that, on the are they engaged in R&D, for example? contrary, FDI spillovers are large and eco- nomically important. Second, there are 6.2 Foreign Direct Investment some results outside the micro productivity Foreign direct investment (FDI) has long literature that do provide evidence for FDI been considered as an important channel for spillovers. technology diffusion. It is a plausible chan- A case study of Intel’s FDI into Costa nel from a theoretical standpoint, because an Rica, for example, already mentioned in influential theory says that firm-specific section 5 above, provides interesting infor- technology is transferred across internation- mation on how widespread the changes can al borders by sharing technology among be that FDI by a major high-technology multinational parents and subsidiaries (see, company can trigger in a relatively small e.g., James Markusen 2002). There are a country (Larrain, Lopez-Calva, and number of models showing how multination- Rodriguez-Claré 2000).59 al enterprises (MNEs) might generate tech- Other authors have provided econometric nological learning externalities for domestic evidence on whether multinational enter- firms, for instance through labor training and prises (MNEs) raise the rate of international turnover (Andrea Fosfuri, Massimo Motta, technology transfer as measured by patent and Thomas Rønde 2001) or through the citations (Steven Globermann, Kokko, and provision of high-quality intermediate inputs Sjohölm 2000; Branstetter 2001a; and Jasjit (Rodriguez-Claré 1996).57 Singh 2003). Here, the results are less clear. Whether FDI indeed generates substan- First of all, MNE subsidiaries could either tial technological externalities for domestic disseminate technology to domestic firms of firms is also a major policy issue, because their host country (inward FDI technology governments all over the world spend large transfer), or MNE subsidiaries might pick up amounts of resources to attract subsidiaries new technologies from the firms in the host of MNEs to their jurisdiction. For instance, country (outward FDI technology sourcing). in 1994, the U.S. state of Alabama spent It turns out that the relative strength of these $230 million, or $150,000 per newly created effects is estimated differently, although the job, to attract a new plant of Mercedes-Benz technology sourcing effect appears to be

56 Industry heterogeneity has recently been empha- 58 In the words of another author, “today’s policy litera- sized in the literature on FDI (see 6.2); Clerides, Lach, ture is filled with extravagant claims about positive and Tybout (1998), however, do generally not estimate spillovers from FDI, [but] the hard evidence is sobering” major differences across industries. (Dani Rodrik 1999, p. 37). 57 See Magnus Blomström and Ari Kokko (1998) and 59 It is an open question whether, or to what extent Kamal Saggi (2000) for a more detailed discussion of pos- these are spillover effects; see also the discussion at the sible spillover channels associated with FDI. end of this section. sp04_Article 2 8/16/04 7:49 AM Page 770

770 Journal of Economic Literature, Vol. XLII (September 2004)

stronger.60 This result—MNE subsidiaries because of heterogeneity across sectors and learn more from the firms in their host coun- across firms. For this reason, the literature try than vice versa—might be indicative of a on FDI spillovers has moved towards using number of problems, however. micro (firm or plant) level data (e.g., Aitken The first is firm heterogeneity: MNE sub- and Harrison 1999; Girma and Wakelin sidiaries are larger and more technologically 2001; Haskel, Pereira, and Slaughter 2001; intensive than the average firm in the host Griffith, Redding, and Simpson 2003; Keller 61 country, and this might be the reason why and Yeaple 2003). These papers present they are good at sourcing technology. This regressions of productivity on FDI and a interpretation seems to be confirmed by number of control variables (they are associ- Singh’s (2003) finding that patent citations ation studies in the sense of section 5.2 between two MNE subsidiaries are stronger above). If productivity growth of domestic than either from MNE subsidiary to a firms is systematically higher in industries domestic firm or the reverse. The second with more FDI than in industries with less issue is endogeneity. It could be that one FDI—and endogeneity and other problems finds MNE subsidiaries to be sourcing more can be ruled out, see section 5.1 above—this technology than they provide because the provides support for the FDI spillovers MNE parent set up the subsidiary with the hypothesis. Irrespective of the precise explicit goal of technology sourcing, while spillover mechanism, a greater presence of the average host-country firm, in contrast, foreign-owned subsidiaries in a particular has not made a comparable location deci- industry is plausibly conducive to domestic sion. This suggests that the estimates are not firms’ technological learning if they belong fully comparable, and future research is to this industry. needed to settle this issue. What is the evidence on this? In Aitken The value of a patent is difficult to esti- and Harrison (1999), the authors estimate a mate (see, e.g., Pakes 1986). An important negative relationship between FDI and pro- issue in the patent citation studies is there- ductivity for a sample of Venezuelan plants. fore how the economic significance of tech- The finding of a negative relationship points nology diffusion measured in this way should to a specification error: while FDI spillovers be assessed. A large literature has thus tried may or may not exist, there is no reason to to estimate directly the extent to which FDI believe that they would be negative. One leads to productivity increases for domestic possibility, raised by Aitken and Harrison firms. Xu (2000), for instance, uses the U.S. (1999), is that their results capture the Bureau of Economic Analysis’ comparable increased competition through FDI in addi- 62 data on U.S. outward FDI into forty coun- tion to spillovers. Haskel, Pereira, and tries over almost thirty years (between 1966 Slaughter (2001), among others, control for and 1994). He finds generally a positive rela- such competition effects. These authors, as tion between FDI and productivity growth, well as Girma and Wakelin (2001) and which is stronger in the richer than in the Griffith, Redding, and Simpson (2003) study poorer countries. inward FDI for the United Kingdom. While Xu’s (2000) analysis is at the manufactur- Girma and Wakelin’s results are somewhat ing level, which may cause aggregation bias mixed, Haskel, Pereira, and Slaughter (2001) and Griffith, Redding, and Simpson 60 Branstetter (2001a) finds evidence for spillovers of both types for Japanese FDI in the U.S., Singh (2003) 61 Görg and Greenaway (2002) is a recent survey. finds the sourcing effect to be stronger than the transfer 62 Another possibility is endogeneity: FDI into effect in a sample of 10 OECD countries, and Globerman, Venezuela targets sectors in which Venezuelan firms are rel- Kokko, and Sjöholm (2000) find only evidence for the atively weak. Aitken and Harrison do not treat endogeneity outward FDI effect. explicitly. sp04_Article 2 8/16/04 7:49 AM Page 771

Keller: International Technology Diffusion 771

(2003) present evidence in support of a sig- in most previous studies, all employees are nificant positive FDI spillover effect. The allocated to one, the so-called major indus- implied economic magnitudes are fairly try. To the extent that MNE subsidiaries are small, however, certainly relative to the sub- diversified and their industry mix changes sidies that have been paid to attract FDI.63 over time, the major-industry method of In contrast, Keller and Yeaple’s (2003) measuring FDI is prone to year-to-year study of recent FDI activity in the United volatility that is essentially noise. Keller and States suggests that FDI spillovers are pos- Yeaple (2003) show that for their sample, itive and large. They estimate that FDI in the preferred FDI data yields estimates that the United States accounts for about 11 are seven times as large as those with the percent of U.S. manufacturing productivity inferior by major-industry FDI data. growth in that period. This estimate is We can summarize the literature on FDI much larger than Haskel, Pereira, and spillovers as follows. In contrast to the earli- Slaughter’s (2001) as well as other esti- er literature, recent micro productivity stud- mates, and the question is why. There seem ies tend to estimate positive, and in some to be two primary reasons. cases also economically large spillovers asso- The first reason is that FDI spillovers ciated with FDI. This difference does not appear to be much stronger in relatively appear to be primarily due to endogeneity high-technology than in relatively low-tech- (or other problems).65 Moreover, although nology sectors.64 This explains in part the current evidence from micro productivi- Keller and Yeaple’s larger FDI spillover ty studies comes from the United Kingdom estimate, because in their Compustat sam- and the United States, there are reasons to ple, the share of high-technology firms is believe that the findings might apply in higher than in the broader sample of other countries as well. If these micro pro- Haskel, Pereira, and Slaughter (2001), for ductivity FDI spillovers hold up in the example. The strong difference between future, it would also provide support for the sectors, which is implicitly present also in FDI learning effects that are found in some Griffith, Redding, and Simpson (2003), of the case studies. table 5—suggests that FDI might be an There are other important questions. First, important conduit for international technol- are there effects of FDI outside the industry ogy diffusion, but only some forms of FDI in which MNE subsidiaries are active? It is that occur in a limited set of sectors. Other plausible to assume that to the extent that FDI, for instance that primarily seeking MNE subsidiaries buy local intermediate lower factor costs, is not likely to generate inputs, they have an incentive to provide major spillovers. Moreover, the practice of technological knowledge to their suppliers; a pooled FDI spillover estimation—across all stronger incentive at any rate than to local manufacturing industries—is likely to lead competitors that produce the same product. to misleading results. Maurice Kugler (2002) as well as Blalock and The second finding of Keller and Yeaple Gertler (2002) have found some evidence for underscores the importance of good meas- such vertical FDI spillovers.66 ures of foreign economic activity. They use Second, could there be more general detailed U.S. statistics to allocate the externalities? Larrain, Lopez-Calva, and employees of a given MNE subsidiary to potentially different industries. In contrast, 65 As noted in section 5, instrumental variable estima- tions are presented in Griffith, Redding, and Simpson 63 Haskel, Pereira, and Slaughter (2001) present a cal- (2003) and Keller and Yeaple (2003). In neither case are culation of costs and benefits. the least-squares estimated FDI spillovers larger than the 64 The two types of sectors are distinguished by R&D corresponding IV-estimates. intensity; see Keller and Yeaple (2003). 66 See also Beata Smarzynska Javorcik (2004). sp04_Article 2 8/16/04 7:49 AM Page 772

772 Journal of Economic Literature, Vol. XLII (September 2004)

Rodriguez-Claré (2000) argue that among import share weights (see section 5).67 He the most important consequences of the finds that within-country spillovers are much chipmaker Intel’s recent FDI into Costa stronger than between-country spillovers. Rica was that (1) Intel funded schools that More evidence for stronger diffusion within taught local workers certain skills (but by than across countries comes from Eaton and no means Intel specific skills), and (2) Kortum’s (1999) study: these authors esti- Intel’s FDI served as a signal for other mate that for the G-5 countries, the rate of potential foreign investors that it might be domestic technology diffusion is about 200 time to invest into Costa Rica now. times the size of the average rate of interna- Estimating such signaling effects (if they tional technology diffusion between the G-5 exist ) is a challenge, because the more countries.68 diffuse and indirect the alleged external- In contrast, Irwin and Klenow (1994) do ities are, the more difficult it becomes to not find stronger within country spillover identify clean causal effects. compared to across country spillovers. Irwin and Klenow estimate that for eight vintages of semiconductors introduced between 1974 7. Geographic Effects on International and 1992, the spillovers from one U.S. firm Technology Diffusion to another U.S. firm are not significantly stronger than those between an U.S. firm Global technology spillovers favor income and a foreign firm. The different results convergence, and local spillovers tend to might be obtained because Irwin and lead to divergence, no matter through Klenow’s spillovers, which are identified which channel technology diffuses. In addi- from the effects of cumulative production on tion to research on channels of diffusion, market shares, are different from knowledge another strand of the literature has exam- spillovers as measured in the other studies.69 ined international technology diffusion in its It could also have to do with the particulars geographic dimension (Jaffe, Trajtenberg, of the semiconductor industry at the time. and Rebecca Henderson 1993; Douglas Most of the relatively small number of firms Irwin and Peter Klenow 1994; Eaton and were located in the United States and in Kortum 1999; Branstetter 2001b; Laura Japan, which means that the scope for iden- Bottazzi and Peri 2000; and Keller 2002a). tifying the within versus between country An advantage of this is that geography is difference is limited. arguably exogenous in this process. One question is whether technology diffu- 67 Patenting data by technological field allows sion within countries is stronger than across Branstetter to compute weights that are increasing in the countries. The evidence generally supports similarity of two firms’ patenting activities; this captures this hypothesis, although there are excep- the idea that R&D expenditures of another firm are more likely to generate spillovers, the closer the two firms are in tions. In particular, Jaffe, Trajtenberg, and technology space. 68 ≠ Henderson (1993) compare the geographic The average international diffusion rate, ni, for n i, in Eaton and Kortum (1999) is about 0.088, while the location of patent citations with that of the = domestic rate of diffusion ( ni, for n i) is fixed at 17.7; see cited patents in the United States. They find their table 2. that U.S. patents are significantly more often 69 Irwin and Klenow (1994) estimate learning-by-doing cited by other U.S. patents than they are spillovers, which are more closely related to human-capital models (e.g. Robert Lucas 1988) than to models of techno- cited by foreign patents. This is confirmed logical change. They might be different in that they meas- by Branstetter (2001b), who uses R&D and ure improvements in the efficiency of an already adopted patenting data on U.S. and Japanese firms to technology, whereas international technology spillovers may capture to a greater extent learning about new and compute weighted R&D spillover stocks not-yet-adopted technologies. Empirical analysis has not analogous to Coe and Helpman’s (1995) been able to clearly separate one from the other so far. sp04_Article 2 8/16/04 7:49 AM Page 773

Keller: International Technology Diffusion 773

The analysis has therefore been extended other changes. Keller (2002a) examines this beyond the national versus international by estimating different decay parameters (δ distinction by estimating spillovers condi- in equation 16) for the late 1970s and the tional on geographic distance and the coun- early 1990s. The estimates indicate that δ tries’ locations relative to each other has shrunk substantially over time in (Keller 2001, 2002a). Keller (2002a) uses absolute value, suggesting that the degree industry-level productivity in nine mostly of localization has become smaller. This is smaller OECD countries to R&D in the G- consistent with the changes mentioned 5 countries (France, Germany, Japan, the above leading to a greater international United Kingdom, and the United States) diffusion of technology. using a simple exponential decay function A major concern is that the estimated in distance: geography effect may be spurious, perhaps   due to unobserved heterogeneity across − lnTFP=+ ln S∑ S × e Dc'c locations. This issue is addressed in a num- cit cit c'it  cG′∈ 5 ber of studies, and while the proposed solu- + ′ + ε tions are clearly imperfect, overall the X.cit (16) results suggest that geography in fact is an

Here, Dc'c is the geographic distance important determinant of technology diffu- between countries c’ and c, and X is a vector sion.70 Another concern seems to be more of control variables. If is estimated to be important. This is that we still need to find greater than zero, variation in productivity is out exactly what the geography effect means best accounted for by giving a lower weight in terms of economic analysis: does this cap- to R&D conducted in countries that are ture trade costs, for instance? We know that located relatively far away, whereas if 0, trade volumes are strongly declining with geographic distance and relative location do distance (Leamer and Levinsohn 1995), so not matter. Keller (2002a) finds that is pos- this seems clearly a possibility. Keller (2001) itive, and moreover, the decay of technology is a first attempt to explain the geographic diffusion implied by the estimate is substan- effects in international technology diffusion tial: with every additional 1,200 kilometers in terms of imports, foreign direct invest- distance there is a 50-percent drop in tech- ment patterns, and person-to-person com- nology diffusion. Applying this estimate to munication, but more work is needed that Australia, for example, with its remote geo- reveals what geography stands for in terms graphic location relative to the G-5 countries, of economic models and behavior. would suggest that Australia benefits extremely little from technology created in 8. Human Capital and R&D as the G-5 countries. Along the same lines, Determinants of International Bottazzi and Peri (2003) find a strong geo- Technology Diffusion graphic decay in their analysis of technology As international technology diffusion diffusion between European regions using a becomes stronger over time, this favors similar framework. These studies suggest income convergence, and vice versa, weaker that technology is highly geographically localized in particular regions and countries. 70 A related question is whether the degree See Jaffe, Trajtenberg, and Henderson (1993), e.g., who adopt a control group approach: they examine cited of localization has fallen in recent years. and not cited patents that are matched in terms of other- This may be expected as a consequence of wise similar characteristics. See also Keller’s (2002a) dis- transport cost improvements, information cussion of whether his results may be explained by geographic distance being correlated with technological and communication technology innova- distance (i.e., that geographically proximate countries pro- tions, increased multinational activity, and duce relatively similar products). sp04_Article 2 8/16/04 7:49 AM Page 774

774 Journal of Economic Literature, Vol. XLII (September 2004)

diffusion makes divergence more likely. There is evidence that human capital facil- However, this result applies only if the trend itates the adoption of new technology in a towards (or away from) greater diffusion is closed economy setting (e.g., Ann Bartel and global, affecting all countries to the same Lichtenberg 1987). Does the same also hold extent. In fact, there seem to be big differ- true for international technology adoption? ences in how effective countries are in adopt- The evidence presented in a number of ing foreign technology. Given that the bulk of papers suggests the affirmative. Eaton and new technology is created in a handful of the Kortum (1996) find that inward technology world’s richest countries, greater technology diffusion is increasing in the level of a coun- diffusion could in fact lead to divergence in try’s human capital. In Francesco Caselli and the world income distribution: this will hap- Wilbur Coleman’s (2001) study, inward tech- pen if today’s richer countries are on average nology diffusion is captured by imports of better at adopting foreign technology than office, computing, and accounting machin- today’s poorer countries. Alternatively, poorer ery, on the grounds that many countries do countries could gain more from better tech- not have a domestic computer industry, so nology diffusion than richer countries. After that computer technology comes necessarily all, poorer countries have more to gain from abroad. They find that computer from it, or, in the words of Alexander imports are positively correlated with human Gerschenkron (1962), latecomers may benefit capital. The results presented by Xu (2000) from their relative backwardness. are consistent with the importance of human This underlines the importance of identi- capital for technology diffusion as well. He fying the major determinants of successful finds that the reason relatively rich countries technology diffusion. Among the many that benefit from hosting U.S. multinational sub- have been proposed, two stand out: human sidiaries while poorer countries do not as capital and R&D expenditures. Both are much has to do with a threshold level of associated with the notion of absorptive human capital in the host country. capacity, the idea that a firm or country Does absorptive capacity in the form of needs to have a certain type of skill in order R&D have a similar effect on technology to be able to successfully adopt foreign diffusion? To date there exist only a few technology; see Keller (1996) for a formal- studies. In one of them, Griffith, Redding, ization of this idea.71 These skills can come, and John Van Reneen (2000) use industry- first of all, in the form of human capital level data from twelve OECD countries for (Richard Nelson and Edmund Phelps the years 1974 to 1990 to study the main 1966). They can also be in the form of R&D determinants of productivity dynamics. spending, a notion that was first emphasized They show that conditional on a certain pro- by Wesley Cohen and Daniel Levinthal ductivity gap to the leader country, subse- (1989). These authors argue that R&D quent productivity growth in an industry is investments are necessary for a firm to higher, the higher are its R&D expenditures. acquire outside technology. This is because This is consistent with R&D playing a role R&D is critical to enabling the firm to similar to that of human capital in providing understand and evaluate new technological the necessary skills for technology adoption. trends and innovations. The second question is whether condi- tional on human capital and R&D speeding 71 Others distinguish specific and currently embodied up the diffusion of technology, there is more from more general skills, and argue that technology diffu- likely to be convergence or divergence. As sion destroys the former skills in the short-run, while in the discussed above, this will depend on the dis- long-run there is a complementary relationship between technical change and skills (Oded Galor and Daniel tribution of human capital and R&D across Tsiddon 1997). countries, and the evidence on this is mixed sp04_Article 2 8/16/04 7:49 AM Page 775

Keller: International Technology Diffusion 775

so far. First, the human-capital threshold the same time geographic localization means identified by Xu (2000) might not have been that d is not equal to zero either. Going surpassed by poor countries, which suggests beyond these results, this section asks what that differences in human capital could lead we know about the relative magnitude of to divergence. Second, a number of FDI domestic and foreign technology sources. spillover studies have found that it is the less Most of the estimates we have come from productive firms that benefit most from for- analyzing samples of relatively rich coun- eign technology (Haskel, Pereira, and tries. In poor countries, there is very little Slaughter 2001; Griffith, Redding and comparable data on domestic technology Simpson 2003; and Keller and Yeaple 2003), investments. The data on the most important and that favors convergence. foreign technology sources that is available— One difference between Xu (2000) on the R&D in the G-7 countries—by itself is thus one side, and the FDI studies on the other not sufficient to assess the relative foreign side, is that the latter do not specifically contribution for poorer countries. In order to focus on the role that a firm’s own R&D (or get some idea nevertheless, I will rely on the skills) might have for convergence. A possi- differences in relative contribution of foreign ble explanation for these seemingly contra- technology among OECD countries. dictory results could be that while the There are several different ways of assess- distribution of R&D and skills in favor of the ing the relative importance of foreign tech- relatively rich is as such a force towards nology diffusion, roughly corresponding to divergence, this is more than outweighed by the different empirical methods described in the stronger benefits that the relatively poor section 5. One approach is to compare the derive from foreign technology that are not TFP elasticities of domestic and foreign conditional on skills and R&D. It is of course R&D. Coe and Helpman (1995) estimate a also possible that convergence of countries domestic R&D elasticity of about 23 percent and convergence of firms in a particular for the relatively large G-7 countries, where- country are not comparable in this way. as the corresponding foreign R&D elasticity More work in this area is clearly needed. is only about 6 percent (these are approxi- mately equal to d and f in equation 12’’). Thus, for the G-7 countries the relative con- 9. Quantifying Domestic and Foreign tribution from foreign R&D is about 21 per- Sources of Productivity Growth cent (0.06 divided by 0.06 plus 0.23). This is Building on the analysis in section 5, the almost the same as in Keller’s (2002b) analy- following equation shows the relation sis of the G-7 countries plus Sweden using

between a country’s productivity (fc) and data at the industry level, where he arrives at f domestic (Sc) as well as foreign R&D (Sc ): a share of about 20 percent for foreign R&D d f f in the total effect on productivity. ln fSS=+ln + ln +ε , (12’’) cc c c c For the smaller OECD countries, Coe Two hypotheses regarding extreme cases of and Helpman estimate a domestic R&D international technology diffusion can be elasticity of about 8 percent and a foreign stated easily. If f is equal to zero, there is no R&D elasticity of about 12 percent. Thus, international technology diffusion at all, the ratio of foreign to domestic effect is esti- whereas if d is equal to zero, then there is mated to be 3:2 for smaller and 1:4 for larg- perfect international technology diffusion, er OECD countries. What lies behind the or, a global pool of technological knowledge. result that foreign R&D is relatively much The discussion so far has indicated that nei- more important for smaller countries? Here, ther of these two cases finds empirical sup- this follows because smaller countries are port—f is clearly greater than zero, but at more open than larger countries—recall sp04_Article 2 8/16/04 7:49 AM Page 776

776 Journal of Economic Literature, Vol. XLII (September 2004)

from section 6.1 that the foreign R&D effect This may suggest that more generally, the is proportional to a country’s overall import smaller a country, the more intensively it share in this specification. relies on domestic technology sources. If so, This openness effect points to country size it would be interesting to find out whether as an important determinant of the relative this is only a country size, or also a distance- importance of foreign versus domestic tech- to-the-frontier phenomenon. The latter nology sources. Is the relative contribution of would mean that even though poor countries foreign R&D indeed so much larger for small receive almost all of their technology from compared with large countries? A first look abroad—there is a vast pool of technology at other evidence seems to suggest yes. For out there—it is domestic technology efforts instance, Eaton and Kortum (1999) estimate that matter most for them. This would be that the part of productivity growth that is consistent with theories where technology due to domestic as opposed to foreign R&D that is invented in frontier countries is less is between 11 percent and 16 percent in appropriate for poorer countries (e.g., Germany, France, and the United Kingdom, Susanto Basu and David Weil 1998). around 35 percent for the larger Japan, and about 60 percent for the United States. 10. Concluding Discussion Keller (2002a) computes the domestic and foreign share from his distance-adjusted esti- What have we learned from this literature mate of effective R&D (equation 16). For so far, and what do we still need to know? I nine countries that are smaller than the will begin by giving a snapshot summary of United Kingdom, Keller estimates that the the evidence in some major areas. The domestic source share is about 10 percent, empirical literature is still quite fragmented, even smaller than Eaton and Kortum’s esti- but there are signs of consolidation which I mate of 11 percent for the United Kingdom. will highlight. I will turn then to an outlook That size matters in some way is clear, on what research might have a particularly however, because in the limit, as a country’s high payoff in the future. share in the world approaches one, the share For most countries, foreign sources of due to foreign technology goes to zero. To technology are of dominant importance (90 control for size differences, one possibility is percent or more) for productivity growth to use GDP data. Doing this gives for the (section 9). This fact underlines the signifi- United States a ratio of domestic technology cance of international technology diffusion share to GDP share of 60 percent to 48 per- and the importance of finding out which cent, or 1.24, whereas the corresponding activities determine its success, and whether value for the Netherlands is 2.18.72 This sug- substantial technology spillovers are associat- gests that relative to GDP, productivity ed with them. Foreign sources of technology growth in the United States relies less on are more important for small (and relatively domestic sources of technology than in the poor) than for bigger countries, which is Netherlands. what one expects given the size difference in domestic R&D investments. 72 These figures are based on results in Eaton and There is no indication, however, that inter- Kortum (1999) and Keller (2002a); GDP data from Heston, national learning is inevitable or automatic. Summers, and Bettina Aten (2002). In the year 1990, the Neither does it seem to be easier for coun- United States contributed about 48 percent to the com- bined size of the G-5 countries plus Sweden, Denmark, and tries that are relatively backward, as the Netherlands, while the Netherlands accounted for 2.2 Gerschenkron (1962) has called it. It could percent. Eaton and Kortum estimate that about 60 percent well be that controlling for size, the poorer a of U.S. productivity growth was due to U.S. R&D, while Keller finds that about 4.7 percent of Dutch productivity country is, the greater is the importance of growth can be attributed to Dutch R&D. domestic technology investments. Instead of sp04_Article 2 8/16/04 7:49 AM Page 777

Keller: International Technology Diffusion 777

simply being far from the frontier, the suc- global. In fact, technology is not global today, cess of countries is in part explained by how as the studies on the localization of technolo- their firms engage in international economic gy diffusion have shown. Yes, we can zap activities. What is the evidence on this so far? technology in the form of computer pro- Early on, international trade has been sug- grams or designs easily around the globe, but gested as a major channel for technology dif- no, this does not mean that there is effective- fusion. With regard to imports, I think that ly a global pool of technology. Perhaps it is overall the evidence now supports the notion because technology is in part tacit, requiring that importing is associated with technology the sender at times to actually go there and spillovers. At the same time, there is still an implement the technology, that there is still a abundance of disparate results that need to geographic pattern to technology diffusion— be brought together and quantified. although the grip of geography has become Moreover, we do not know yet how strong weaker recently. We still need to know more diffusion is through embodied technology about why geography matters—explaining in intermediate goods versus other tech- changes in the geographic pattern of technol- nology diffusion associated with imports. ogy diffusion will tell us much, I think, about Learning effects from exporting have been the key mechanisms, especially in analyses found in the case study literature, whereas that include both rich and poor countries. authors of econometric studies take a much In order to decide what kind of research more skeptical view. While the econometric would be most useful at this point, it is evidence on learning-by-exporting effects is worth recalling what the options are. not as clear cut and negative as it is often Broadly speaking, those are case studies, stated, right now I do not see evidence in econometric analyses, and simulation stud- favor of strong important learning-by- ies. As long as their respective strengths and exporting effects either. weaknesses are recognized, all three can What about learning effects for domestic potentially prove to be very useful. Of great firms from FDI? The FDI literature seems importance is whether the data that is to be closer to a consensus than the trade employed is closely related to issues of tech- literature is right now. For instance, both nology and technology diffusion. For micro-econometric studies and case studies instance, FDI spillovers estimated from point in the same direction. The evidence data on foreign subsidiaries’ (and their par- suggests that there can be FDI spillovers, ents’) R&D should tell us much more on but they do not occur everywhere to the technology transfer than a variable like the same degree. The remaining questions foreign share of employment. The general include the following: how large are vertical principle is: if you know something about compared to horizontal spillovers, and actual technological activity, try to use it. which firms—stronger or weaker—benefit An important question is the appropriate most from spillovers. In addition, we need level of data aggregation. This issue was to know whether FDI spillovers in richer moot for a long time in international eco- and poorer countries are equally strong, nomics because there was very little micro and last not least whether they are quantita- data. Recently however, micro data sets have tively large enough to justify the large sub- started to become available. It is thus impor- sidies that governments provide to attract tant to know the differences between micro multinationals. and more aggregate data for studying inter- There is a thin line between recognizing national technology diffusion, and which of the importance of international technology the two, if any, is preferred. Not only can diffusion and getting carried away in the one use micro data today, but moreover it belief that in today’s world technology is seems to matter a great deal for the results: sp04_Article 2 8/16/04 7:49 AM Page 778

778 Journal of Economic Literature, Vol. XLII (September 2004)

generally, the higher is the level of aggrega- better estimation of micro behavior, as the tion, the stronger tends to be the evidence data is recorded right at the decision-taking for externalities and learning effects. What level. And to reach their full potential, can explain this difference between micro empirical micro studies of technology diffu- and more aggregate evidence? sion will probably involve modeling micro Micro data can capture heterogeneity structure as well, because it is the structure across firms, something that features promi- that shapes behavior. As of now, there is lit- nently even in narrowly defined industries, tle of that. For instance, currently, empirical whereas industry- and aggregate-level stud- studies of technology diffusion are not based ies cannot control for this and may suffer on models whose structure explicitly allows from composition and aggregation biases for learning and spillover externalities that tend to lead to inflated spillover esti- between firms. That is, so far the micro evi- mates. I think that this is indeed sometimes dence is based on “no-externality” models. the case. At the same time, firm heterogene- This suggests that the micro evidence may ity is less important if one is primarily inter- become stronger once externalities are ested in industry-level outcomes. For included in the models as a matter of theory example, it may not matter too much for (some work in this direction is Clerides, economic policy whether industry produc- Keller, and Tybout 2003). tivity rises because all firms benefit from If the idea of differencing is to control for spillovers or because only the more produc- the averse effects of heterogeneity (across tive firms benefit and become larger while firms or over time), where does one stop? Of the least productive firms exit. course, solving the heterogeneity issue fun- Simultaneity and endogeneity seem to be damentally means conducting a case study, more important issues than aggregation, and with a sample of one observation—but it is in this respect there is little difference not necessarily advisable to go that far, at between micro and more aggregate studies: least not as the sole empirical methodology. for instance, interpreting a cross-sectional More generally, can there be a problem of correlation of foreign ownership and pro- overdifferencing? An important issue is that ductivity as evidence for FDI spillovers technology data tends to be not of very high would be just as inappropriate at the firm quality, and this limits the amount of differ- level as it is at the aggregate level. Instead, encing that one can do. Once the estimated the work in question is more or less impor- coefficients look like a collection of random tant depending on how thorough the authors numbers scattered around zero with typical- are in trying to identify a truly causal effect. ly large standard errors, the researcher has Most strategies for doing that rely on com- probably asked too much of the data. paring sets of firms. This comparison, or, tak- Similarly, is it possible to probe so deeply ing differences, may allow us to see whether into the structure of micro behavior that one the treatment (say, inward FDI) has an effect. fails to notice more diffuse, more aggregate, There is no doubt that differencing is crucial and slower-acting effects? As noted above, for capturing a causal effect. At the same one benefit of attracting a major foreign time, the assumptions on the nature of the investor, according to Larrain, Lopez-Calva, differences are often quite simple (e.g., firm- and Rodriguez-Claré (2000), is that it sends specific time-invariant heterogeneity), and it a signal to other potential foreign investors. is not clear whether the results would remain It is difficult to see how a deeper analysis of the same if firms are actually allowed to react micro structure would help to detect these to a major change in their environment. types of effects. More generally, I expect the biggest con- Summarizing, I think that we can expect tribution of micro data sets to come from a to obtain additional major results on interna- sp04_Article 2 8/16/04 7:49 AM Page 779

Keller: International Technology Diffusion 779

tional technology diffusion in the future, and of high-tech goods. Rather, the goods char- this will be in part because of the increased acteristics are only part of what is important. usage of micro data and estimation. At the India, for example, aimed at producing rela- same time, it will be important to keep an tively advanced products during its era of open mind and not expect too much from import-substitution policy, but their quality any one empirical methodology. was low compared to international stan- What are the policy implications of this lit- dards. In contrast, Southeast Asian countries erature? The evidence is not strong enough such as Hong Kong and South Korea were yet to provide support for specific policy initially specializing in relatively low-tech measures, such as a particular subsidy to a products, and moved slowly but successfully multinational enterprise for locating in a into the range of higher-tech products. country. This would require more agreement Instead, technological knowledge on quantitative effects than there is right now. spillovers appear to be resulting from a At this point, the results suggest that the deliberate commitment to learning and international dimension of technological matching international performance stan- change is of key importance for most coun- dards through ongoing interaction with for- tries. In this situation, a closed-off internation- eigners. Local efforts are clearly necessary al economic regime must have detrimental for successful technology adoption. At the welfare consequences for a country. same time, the ongoing interaction with for- There is also evidence that the relative eign firms and consumers seems to be a importance of international technology diffu- process of knowledge discovery for firms sion has been increasing with the level of eco- that cannot be had from interacting only nomic integration in the world. This suggests with other domestic firms. To model as well that the performance advantage of outward- as empirically capture these in a convincing oriented economies over inward-oriented way continues to be a challenge. economies will be higher in the future than it is today. What if future research were to REFERENCES establish that international technology diffu- Abramovitz, Moses. 1956. “Resource and Output Trends in the United States since 1870,” Amer. Econ. sion raises productivity more in advanced Rev. 46:2, pp. 5–23. than in less-developed countries? The impli- Acemoglu, Daron; Simon Johnson and James cation for less-developed countries would not Robinson. 2001. “The Colonial Origins of Comparative Development,” Amer. Econ. Rev. 91:5, be that a turn towards autarky is beneficial— pp. 1369–401. on the contrary, given the evidence on posi- Aghion, Philippe and Peter Howitt. 1992. “A Model of tive productivity effects from international Growth through Creative Destruction,” Econometrica 60, pp. 323–51. technology diffusion. Rather, if the benefits ———. 1998. Endogenous Growth Theory. from operating in an international economic Cambridge, MA: MIT Press. environment differ across countries, this sug- Aitken, Brian and Ann Harrison. 1999. “Do Domestic Firms Benefit from Foreign Direct Investment? gests we should investigate further the major Evidence from Venezuela,” Amer. Econ. Rev. 89:3, reasons for this—what works, and why? pp. 605–18. While there is no consensus yet on the Arrow, Kenneth. 1969. “Classificatory Notes on the Production and Transmission of Technological exact magnitude of spillover benefits, it is Knowledge,” Amer. Econ. Rev. Pap. Proceed. 59:2, clear that well-functioning markets and an pp. 29–35. undistorted trade and foreign-investment Baldwin, Richard. 1989. “Sunk Costs Hysteresis,” NBER work. pap. 2911, Cambridge, MA. regime are conducive to these learning Banerjee, Abhijit and Lakshmi Iyer. 2003. “History, effects. However, the evidence suggests that Institutions, and Economic Performance: The the latter are quite difficult to pin down. For Legacy of Colonial Land Tenure Systems in India,” BREAD work. pap. one, it does not seem to be primarily a mat- Bartel, Ann and Frank Lichtenberg. 1987. “The ter of simply specializing in the production Comparative Advantage of Educated Workers in sp04_Article 2 8/16/04 7:49 AM Page 780

780 Journal of Economic Literature, Vol. XLII (September 2004)

Implementing Technology,” Rev. Econ. Statist. 69:1, “Is Learning by Exporting Important? Micro- pp. 1–11. dynamic Evidence from Colombia, Mexico, and Bartelsman, Eric and Mark Doms. 2000. Morocco,” Quart. J. Econ. 113, pp. 903–48. “Understanding Productivity: Lessons from Coe, David and Elhanan Helpman. 1995. Longitudinal Microdata,” J. Econ. Lit. 38:3, pp. “International R&D Spillovers,” Europ. Econ. Rev. 569–94. 39:5, pp. 859–87. Basu, Susanto and David Weil. 1998. “Appropriate Coe, David; Elhanan Helpman and Alexander Technology and Growth,” Quart. J. Econ. 113:4, pp. Hoffmaister. 1997. “North-South Spillovers,” Econ. 1025–54. J. 107, pp. 134–49. Bernard, Andrew and Bradford Jensen. 1999. Coe, David and Alexander Hoffmaister. 1999. “Are “Exceptional Exporter Performance: Cause, Effect, There International R&D Spillovers among or Both?” J. Int. Econ. 47:1, pp. 1–25. Randomly Matched Trade Partners? A Response to Bernard, Andrew; Jonathan Eaton, Bradford Jensen Keller,” IMF work. pap. 99/18. and Samuel Kortum. 2003. “Plants and Productivity Cohen, Wesley and Daniel Levinthal. 1989. in International Trade,” Amer. Econ. Rev. 93:4, pp. “Innovation and Learning: The Two Faces of R&D,” 1268–90. Econ. J. 99:397, pp. 569–96. Blalock, Garrick and Paul Gertler. 2002. “Technology David, Paul. 1992. “Knowledge, Property, and the Diffusion from Foreign Direct Investment through Systems Dynamics of Technological Change,” in Supply Chains,” work pap., Haas School Bus. UC Proceedings of the World Bank Annual Conference Berkeley. on Development Economics 1992, Larry Summers Blomström, Magnus and Ari Kokko. 1998. and Anwar Shah, eds. pp. 215–48. “Multinational Corporations and Spillovers,” J. Econ. Dixit, Avinash. 1989. “Entry and Exit Decisions under Surveys 12:3, pp. 247–77. Uncertainty,” J. Polit. Econ. 97:3, pp. 620–38. Blundell, Richard, and Stephen Bond. 1999. “GMM Dornbusch, Rüdiger; Stanley Fischer and Paul Estimation with Persistent Panel Data: An Samuelson. 1977. “Comparative Advantage, Trade, Application to Production Functions,” work. pap. and Payments in a Ricardian Model with a 99/4, Inst. Fiscal Studies, London. Continuum of Goods,” Amer. Econ. Rev. 67:5, pp. Bottazzi, Laura and Giovanni Peri. 2003. “Innovation, 823–39. Demand, and Knowledge Spillovers: Evidence from Easterly, William and Ross Levine. 2001. “It’s Not European Patent Data,” Europ. Econ. Rev. 47, pp. Factor Accumulation: Stylized Facts and Growth 687–710. Models,” World Bank Econ. Rev. 15:2, pp. 177–219. Brainard, Lael. 1997. “An Empirical Assessment of the Eaton, Jonathan; Eva Gutierrez and Samuel Kortum. Proximity-Concentration Trade-off between 1998. “European Technology Policy,” Econ. Pol. 27, Multinational Sales and Trade,” Amer. Econ. Rev. pp. 405–38. 87:4, pp. 520–44. Eaton, Jonathan and Samuel Kortum 1996. “Trade in Branstetter, Lee. 2001a. “Is Foreign Direct Ideas: Patenting and Productivity in the OECD,” J. Investment a Channel of Knowledge Spillovers? Int. Econ. 40:3-4, pp. 251–78. Evidence from Japan’s FDI in the United States,” ———. 1997. “Engines of Growth: Domestic and NBER work. pap. 8015. Foreign Sources of Innovation,” Japan World Econ. ———. 2001b. “Are Knowledge Spillovers 9, pp. 235–59. International or Intranational in Scope? ———. 1999. “International Patenting and Microeconometric Evidence from the U.S. and Technology Diffusion: Theory and Measurement,” Japan,” J. Int. Econ. 53, pp. 53–79. Int. Econ. Rev. 40, pp. 537–70. Caballero, Ricardo and Adam Jaffe. 1993. “How High ———. 2001. “Trade in Capital Goods,” Europ. Econ. Are the Giant’s Shoulders? An Empirical Assessment Rev. 45:7, pp. 1195–235. of Knowledge Spillovers and Creative Destruction in ———. 2002. “Technology, Geography, and Trade,” a Model of Growth,” NBER Macroeconomic Econometrica 70:5, pp. 1741–79. Annual, vol. 8. Olivier Blanchard and Stanley Edmonds, Chris. 2001. “Some Panel Cointegration Fischer, eds. Models of International R&D Spillovers,” J. Caselli, Francesco and Wilbur Coleman II. 2001. Macroecon. 23, pp. 241–60. “Cross-Country Technology Diffusion: The Case of Ethier, William. 1986. “The Multinational Firm,” Computers,” Amer. Econ. Rev. Pap. Proceed. 91:2, Quart. J. Econ. 101, pp. 805–33. pp. 328–35. ———. 1982. “National and International Returns to Caves, Richard; Laurits Christensen and Erwin Scale in the Modern Theory of International Trade,” Diewert. 1982. “Multilateral Comparisons of Amer. Econ. Rev. 72, pp. 389–405. Output, Input, and Productivity Using Superlative Evenson, Robert and Larry Westphal. 1995. Index Numbers,” Econ. J. 92:365, pp. 73–86. “Technology Change and Technology Strategy,” in CIC. 2003. The International Comparison of Prices Handbook of Development Economics, vol. 3A. Jere Program. Center Int. Comparisons, U. Penn. Behrman and T.N. Srinivasan, eds. North-Holland. Clerides, Sofronis; Wolfgang Keller and James Tybout. Fagerberg, Jan. 1994. “Technology and International 2003. “Foreign Direct Investment, Exports, and Differences in Growth Rates,” J. Econ. Lit. 32:3, pp. Spillovers,” in progress, Brown U. 1147–75. Clerides, Sofronis; Saul Lach and James Tybout. 1998. Feldman, Maryann and Frank Lichtenberg. 1997. “The sp04_Article 2 8/16/04 7:49 AM Page 781

Keller: International Technology Diffusion 781

Impact and Organization of Publicly-Funded Hallward-Driemeier, Mary; Giuseppe Iarossi and Research and Development in the European Kenneth Sokoloff. 2002. “Exports and Manufacturing Community,” NBER work. pap. 6040. Productivity in East Asia: A Comparative Analysis Fosfuri, Andrea; Massimo Motta and Thomas Rønde. with Firm-Level Data,” work. pap., UCLA. 2001. “Foreign Direct Investment and Spillovers Hanson, Gordon. 2001. “Should Countries Promote through Workers’ Mobility,” J. Int. Econ. 53, pp. Foreign Direct Investment?” G-24, UN Discus. Pap. 205–22. Series, NY and Geneva. Galor, Oded and Daniel Tsiddon. 1997. “Technological Haskel, Jonathan; Sonia Pereira and Matthew Progress, Mobility, and Economic Growth,” Amer. Slaughter. 2001. “Does Inward Foreign Direct Econ. Rev. 87, pp. 363–82. Investment Boost the Productivity of Domestic Gerschenkron, Alexander. 1962. Economic Firms?” pap. for NBER Summer Institute. Backwardness in Historical Perspective. Cambridge, Heston, Alan; Robert Summers and Bettina Aten. MA: Harvard U. Press. 2002. Penn World Tables Version 6.1, Center Int. Girma, Sourafel and Katharina Wakelin. 2001. Comparisons, U. Penn. “Regional Underdevelopment: Is FDI the Solution? Howitt, Peter. 2000. “Endogenous Growth and Cross- A Semi-Parametric Analysis,” GEP Research Pap. Country Income Differences,” Amer. Econ Rev. 2001/11, U. Nottingham. 90:4, pp. 829–46. Görg, Holger, and David Greenaway. 2002. “Much Ado Hulten, Charles. 2000. “Total Factor Productivity: A About Nothing? Do Domestic Firms Really Benefit Short Biography,” NBER work. pap. 7471. from Foreign Direct Investment?” work pap. U. Irwin, Douglas and Peter Klenow. 1994. “Learning Nottingham. Spillovers in the Semi-Conductor Industry,” J. Polit. Globerman, Steven; Ari Kokko and Frederic Sjöholm. Econ. 102, pp. 1200–27. 2000. “International Technology Diffusion: Evidence Jaffe, Adam and Manuel Trajtenberg. 2002. Patents, from Swedish Patent Data,” Kyklos 53, pp. 17–38. Citations and Innovations: A Window on the Griffith, Rachel; Stephen Redding and Helen Simpson. Knowledge Economy. Cambridge: MIT Press. 2003. “Productivity Convergence and Foreign Jaffe, Adam; Manuel Trajtenberg and Rebecca Ownership at the Establishment Level,” CEPR Henderson. 1993. “Geographic Localization of work. pap. 3765. Knowledge Spillovers as Evidenced by Patent Griffith, Rachel; Stephen Redding, and John Van Citations,” Quart. J. Econ. 108:3, pp. 577–98. Reenen. 2000. “Mapping Two Faces of R&D: Jones, Charles. 1995. “R&D-Based Models of Productivity Growth in a Panel of OECD Growth,” J. Polit. Econ. 103:4, pp. 759–84. Industries,” forthcoming, Rev. Econ. Statist. Katayama, Hajime; Shihua Lu and James Tybout. 2003. Griliches, Zvi. 1979. “Issues in Assessing the “Why Plant-Level Productivity Studies Are Often Contribution of Research and Development to Misleading, and an Alternative Approach to Productivity Growth,” Bell J. Econ. 10, pp. 92–116. Inference,” work pap. Penn State U. ———. ed. 1984. R&D, Patents, and Productivity. U. Keller, Wolfgang. 1996. “Absorptive Capacity: On the Chicago for NBER. Creation and Acquisition of Technology in ———. 1990. “Patent Statistics as Economic Development,” J. Devel. Econ. 49, pp. 199–227. Indicators: A Survey,” J. Econ. Lit. 28:3, pp. 1661–707. ———. 1997. “How Trade Patterns and Technology ———. 1995. “R&D and Productivity: Econometric Flows Affect Productivity Growth,” World Bank Results and Measurement Issues,” in Handbook of Policy Research work. pap. 1831. the Economics of Innovation and Technological ———. 1998. “Are International R&D Spillovers Change. Paul Stoneman ed. , Oxford: Blackwell, pp. Trade Related? Analyzing Spillovers among 52–89. Randomly Matched Trade Partners,” Europ. Econ. Griliches, Zvi, and Jerry Hausman. 1986. “Errors in Rev. 42, pp. 1469–81. Variables in Panel Data,” J. Econometrics 31, pp. ———. 2000. “Do Trade Patterns and Technology 93–118. Flows Affect Productivity Growth?” World Bank Griliches, Zvi, and Jacques Mairesse. 1998. Econ. Rev. 14, pp. 17–47. “Production Functions: The Search for ———. 2001. “Knowledge Spillovers at the World’s Identification,” in Econometrics and Economic Technology Frontier,” CEPR work. pap. 2815. Theory in the 20th Century: Ragnar Frisch ———. 2002a. “Geographic Localization of Centennial Symposium. Steinar Strøm, ed. International Technology Diffusion,” Amer. Econ. Cambridge: Cambridge U. Press, pp. 169–203. Rev. 92, pp. 120–42. Grossman, Gene and Elhanan Helpman. 1991. ———. 2002b. “Trade and the Transmission of Innovation and Growth in the World Economy, Technology,” J. Econ. Growth: 7, pp. 5–24. Cambridge, MA: MIT Press. Keller, Wolfgang and Stephen Yeaple. 2003. ———. 1995. “Technology and Trade,” in Handbook “Multinational Enterprises, International Trade, and of International Economics vol.3. Gene Grossman Productivity Growth: Firm Level Evidence from the and Kenneth Rogoff, eds. North-Holland. United States,” IMF work. pap. 248. Hall, Robert and Charles Jones. 1999. “Why Do Some Klette, Tor and Zvi Griliches. 1996. “The Inconsistency Countries Produce So Much More Output Per of Common Scale Estimators When Output Prices Worker Than Others?” Quart. J. Econ. 114:1, pp. Are Unobserved and Endogenous,” J. Appl. 83-116. Econometrics 11, pp. 343–61. sp04_Article 2 8/16/04 7:49 AM Page 782

782 Journal of Economic Literature, Vol. XLII (September 2004)

Krugman, Paul. 1989. Exchange Rate Instability. ———. 2001b. “Demand-Driven Knowledge Clusters Cambridge, MA: MIT Press. in a Weightless Economy,” work pap., London Kugler, Maurice. 2002. “The Diffusion of Externalities School Econ. from Foreign Direct Investment: Theory Ahead of Rhee, Yong-Whee; Bruce Ross-Larson and Garry Measurement,” work pap. U. Southhampton. Pursell. 1984. Korea’s Competitive Edge: Managing Larrain, B. Felipe; Luis Lopez-Calva, and Andres Entry into World Markets. Baltimore: Johns Hopkins Rodriguez-Claré. 2000. “Intel: A Case Study of U. Press for World Bank. Foreign Direct Investment in Central America,” Rivera-Batiz, Luis and Paul Romer. 1991. “Economic work. pap. 58, Center Int. Develop., Harvard U. Integration and Endogenous Growth,” Quart. J. Leamer, Edward and James Levinsohn. 1995. Econ. 106:2, pp. 531–55. “International Trade theory: The Evidence,” in Rodriguez-Claré, Andres. 1996. “Multinationals, Handbook of International Economics, vol. 3. G. Linkages, and Economic Development,” Amer. Grossman and K. Rogoff, eds. Elsevier: Amsterdam. Econ. Rev. 86, pp. 852–73. Lichtenberg, Frank. 1993. “R&D Investment and Rodrik, Dani. 1999. “The New Global Economy and International Productivity Differences,” in Economic Developing Countries: Making Openness Work,” Growth in the World Economy. Symposium 1992. Policy Essay 24, Overseas Develop. Council, Horst Siebert, ed. Tübingen: Mohr. Washington DC. Lichtenberg, Frank and Bruno van Pottelsberghe de la Romer, Paul. 1990. “Endogenous Technological Potterie. 2000. “Does Foreign Direct Investment Change,” J. Polit. Econ. 98:5, pp. S71–S102. Transfer Technology Across Borders?” forthcoming, Saggi, Kamal. 2000. “Trade, Foreign Direct Rev. Econ. Statist. Investment, and International Technology Transfer: Lucas, Robert. 1988. “On the Mechanics of Economic A Survey,” forthcoming in World Bank Res. Development,” J. Monet. Econ. 22, pp. 3–42. Observer. Lumenga-Neso, Olivier; Marcelo Olarreaga and Scherer, Frederic. 1984. “Using Linked Patent and Maurice Schiff. 2001. “On ‘Indirect’ Trade-Related R&D Data to Measure Interindustry Technology R&D spillovers,” mimeo, World Bank. Flows,” in R&D, Patents, and Productivity. Zvi Maddala, G.S. and In-Moo Kim. 1998. Unit Roots, Griliches ed. U. Chicago for NBER, pp. 417–61. Cointegration, and Structural Change. Cambridge: Segerstrom, Paul; T.C.A. Anant and Elias Dinopoulos. Cambridge U. Press. 1990. “A Schumpeterian Model of the Product Life Markusen, James. 2002. Multinational Firms and the Cycle,” Amer. Econ. Rev. 80:4, pp. 1077–92. Theory of International Trade. Cambridge: MIT Shiue, Carol H. and Wolfgang Keller. 2004. “Markets Press. in China and Europe on the Eve of the Industrial Nelson, Richard and Edmunds Phelps. 1966. Revolution,” CEPR discus. pap. 4420. “Investment in Humans, Technological Diffusion Singh, Jasjit. 2003. “Knowledge Diffusion and and Economic Growth,” Amer. Econ. Rev. 56:2, pp. Multinational Firms: Evidence using Patent Citation 69–75. Data,” work. pap., Grad. School Bus. Admin. and NSF. 2004. “US–China R&D Linkages: Direct Econ. Dept., Harvard U. Investment and Industrial Alliances in the 1990s,” Sjöholm, Frederic. 1996. “International Transfer of NSF Info. Brief NSF 04-236. Knowledge: The Role of International Trade and OECD 2002. Frascati Manual 2002: Proposed Geographic Proximity,” Weltwirtschaftliches Archiv Standard Practice for Surveys on Research and 132, pp. 97–115. Development. Paris: OECD. Teece, David. 1977. “Technology Transfer by ———. 2003. Main Science and Technology Multinational Firms: The Resource Cost of Indicators. Paris: OECD. Transferring Technological Know-How,” Econ. J. 87, Olley, Steven and Ariel Pakes. 1996. “The Dynamics pp. 242–61. of Productivity in the Telecommunications von Hippel, Eric. 1994. “‘Sticky Information’ and the Equipment Industry,” Econometrica 64:6, pp. Locus of Problem Solving: Implications for 1263–97. Innovation,” Manage. Sci. 40:4, pp. 429–39. Pakes, Ariel. 1986. “Patents as Options: Some Xu, Bin. 2000. “Multinational Enterprises, Technology Estimates of the Value of Holding European Patent Diffusion, and Host Country Productivity Growth,” Stocks,” Econometrica 54:4, pp. 755–84. J. Devel. Econ., 62:2, pp. 477–93. Peri, Giovanni. 2002. “Knowledge Flows and Xu, Bin and Jianmao Wang. 1999. “Capital Goods Innovation,” work. pap., UC Davis. Trade and R&D Spillovers in the OECD,” Can. J. Polanyi, Michael. 1958. Personal Knowledge: Towards a Econ. 32:5, pp. 1258–74. Post-Critical Philosophy. Chicago: U. Chicago Press. WIPO. 2003. World Intellectual Property Prescott, Edward. 1998. “Needed: A Theory of Total Organization, www.wipo.org. Factor Productivity,” Int. Econ. Rev. 39:3, pp. 525–51. World Bank. 1999. Knowledge for Development. World Quah, Danny. 2001a. “Technology Dissemination and Development Report 1998/99, Washington DC: Economic Growth,” work. pap., London School Econ. World Bank.