WPS8027 Policy Research Working Paper 8027 Heterogeneous Technology Diffusion and Ricardian Trade Patterns William R. Kerr Development Economics Vice Presidency Operations and Strategy Team April 2017 Policy Research Working Paper 8027 Abstract Migration and trade are often linked through ethnic net- development of new technologies across US cities with his- works boosting bilateral trade. This study uses migration torical settlement patterns for migrants from countries. The to quantify the importance of Ricardian technology dif- instrumented elasticity of export growth on the intensive ferences for international trade. The framework provides margin with respect to the exporter’s productivity growth the first panel estimates connecting country-industry is between 1.6 and 2.4, depending upon weighting. This productivity and exports, and the study exploits heteroge- provides an important contribution to the trade literature neous technology diffusion from immigrant communities of Ricardian advantages, and it establishes a connection of in the United States for identification. The latter instru- migration to home country exports beyond bilateral networks. ments are developed by combining panel variation on the This paper is a product of the Operations and Strategy Team, Development Economics Vice Presidency. It is part of a larger effort by the World Bank to provide open access to its research and make a contribution to development policy discussions around the world. Policy Research Working Papers are also posted on the Web at http://econ.worldbank.org. The author may be contacted at wkerr@hbs.edu@worldbank.org. The Policy Research Working Paper Series disseminates the findings of work in progress to encourage the exchange of ideas about development issues. An objective of the series is to get the findings out quickly, even if the presentations are less than fully polished. The papers carry the names of the authors and should be cited accordingly. The findings, interpretations, and conclusions expressed in this paper are entirely those of the authors. They do not necessarily represent the views of the International Bank for Reconstruction and Development/World Bank and its affiliated organizations, or those of the Executive Directors of the World Bank or the governments they represent. Produced by the Research Support Team Heterogeneous Technology Diffusion and Ricardian Trade Patterns William R. Kerr *† JEL Classification: F11, F14, F15, F22, J44, J61, L14, O31, O33, O57 Key Words: Trade, Exports, Comparative Advantage, Technological Transfer, Patents, Innovation, Research and Development, Immigration, Networks * William Kerr is Dimitri V. D’Arbeloff—MBA Class of 1955 Professor of Business Administration, Harvard Business School, Boston, Massachusetts, and Faculty Research Associate, National Bureau of Economic Research, Cambridge, Massachusetts. His email address is wkerr@hbs.edu. † I am grateful to Daron Acemoglu, Pol Antras, David Autor, Dany Bahar, Nick Bloom, Ricardo Caballero, Arnaud Costinot, Julian Di Giovanni, Robert Feenstra, Fritz Foley, Richard Freeman, Gordon Hansen, Sam Kortum, Ashley Lester, Matt Mitchell, Peter Morrow, Ramana Nanda, Giovanni Peri, Hillel Rapoport, Ariell Reshef, Tim Simcoe, Antonio Spilimbergo, Scott Stern, Sarah Turner, and John Van Reneen for advice on this project and to seminar participants at the eighth AFD-World Bank Migration and Development Conference, American Economic Association meetings, Clemson University, Columbia University, European Regional Science Association meetings, Georgetown University, Harvard University, International Monetary Fund, London School of Economics, MIT Economics, MIT Sloan, NBER High Skilled Immigration Conference, NBER Productivity, Queens University, University of California Davis, University of Helsinki, University of Toronto Rotman, World Bank, and Yale University for helpful comments. This paper is a substantial revision of chapter 2 of my Ph.D. dissertation (Kerr 2005). This research is supported by the National Science Foundation, MIT George Schultz Fund, HBS Research, and the Innovation Policy and Economy Group. Trade among countries due to technology differences is a core principle in international economics. Countries with heterogeneous technologies focus on producing goods in which they have comparative advantages; subsequent exchanges afford higher standards of living than are possible in isolation. This Ricardian finding is the first lesson in most undergraduate courses on trade, and it undergirds many modeling frameworks on which recent theoretical advances build (e.g., Dornbusch et al. 1977, Eaton and Kortum 2002, Costinot et al. 2012). In response to Stanislaw Ulam’s challenge to name a true and nontrivial theory in social sciences, Paul Samuelson chose this principle of comparative advantage due to technology differences. While empirical tests date back to David Ricardo (1817), quantifying technology differences across countries and industries is extremely difficult. Even when observable proxies for latent technology differences are developed (e.g., labor productivity, industrial specialization), cross-sectional analyses risk confounding heterogeneous technologies with other country-industry determinants of trade. Panel data models can further remove time-invariant characteristics (e.g., distances, colonial histories) and afford explicit controls of time-varying determinants (e.g., factor accumulation, economic development, trading blocs). Quantifying the dynamics of uneven technology advancement across countries is an even more challenging task, however, and whether identified relationships represent causal linkages remains a concern. These limitations are particularly acute for developing and emerging economies. This is unfortunate as non- OECD economies have experienced some of the more dramatic changes in technology sets and manufacturing trade over the last thirty years, providing a useful laboratory for quantifying Ricardian effects. This study contributes to the empirical trade literature on Ricardian advantages in three ways. First, it utilizes a panel dataset that includes many countries at various development stages (e.g., Bolivia, France, South Africa), a large group of focused manufacturing industries, and an extended time frame. The 1975–2000 World Trade Flows (WTF) database provides export data for each bilateral route (exporter-importer- industry-year), and data from the United Nations Industrial Development Organization (UNIDO) provide labor productivity estimates. The developed data platform includes substantially more variation in trade and productivity differences across countries than previously feasible. The second contribution is to provide panel estimates of the elasticity of export growth with respect to productivity development. Following the theoretical work of Costinot et al. (2012) that is discussed below, estimations include fixed effects for importer-industry-year and exporter-importer-year. The importer-industry-year fixed effects control, for example, for trade barriers in each importing country by industry segment while the exporter-importer-year fixed effects control for the overall levels of trade between countries (e.g., the gravity model), labor cost structures in the exporter, and similar. While these controls account for overall trade and technology levels by country, permanent differences in the levels of these variables across industries within a country are used for identification in most applications of this approach. This paper is the first to quantify Ricardian elasticities when further modeling cross-sectional fixed effects for exporter-importer-industry observations. This panel approach only exploits variation within industry-level bilateral trading routes, providing a substantially stronger empirical test of the theory. The third and most important contribution is to provide instruments for the labor productivity development in exporting countries. Instruments are essential in this setting due to typical concerns: omitted variable biases for the labor productivity measure, reverse causality, and the potential for significant measurement error regarding the productivity differences across countries. The instruments exploit heterogeneous technology diffusion from past migrant communities in the United States for identification. These instruments are developed by combining panel variation on the development of new technologies across US cities during the 1975–2000 period with historical settlement patterns for migrants and their ancestors from countries that are recorded in the 1980 Census of Populations. The foundation for these instruments is the modeling of Ricardian advantages through differences across countries in their access to the US technology frontier. Recent research emphasizes the importance of immigrants in frontier economies for the diffusion of technologies to their home countries (e.g., Saxenian 2002, 2006, Kerr 2008). These global connections and networks facilitate the transfer of both codified and tacit details of new innovations, and Kerr (2008) finds foreign countries realize manufacturing gains from stronger scientific integration, especially with respect to computer-oriented technologies. Multiple studies document specific channels sitting behind this heterogeneous diffusion. 1 As invention is disproportionately concentrated in the United States, these ethnic networks significantly influence technology opportunity sets in the short-run for following economies. This study uses heterogeneous technology diffusion from the United States to better quantify the importance of technology differences across countries in explaining trade patterns. Trade between the United States and foreign countries is excluded throughout this study due to network effects operating alongside technology transfers. Attention is instead placed on how differential technology transfer from the United States—particularly its industry-level variation by country—influences exports from the foreign country to other nations. Said differently, the study quantifies the extent to which India’s exports, for example, grow faster in industries where technology transfer from the United States to India is particularly strong. This provides an important complement in the migration literature to the typical focus on how ethnic networks boost bilateral trade. The instrumented elasticity of export growth on the intensive margin with respect to the exporter’s productivity growth is 2.4 in unweighted estimations. The elasticity is 1.6 when using sample weights that interact worldwide trade volumes for exporters and importers in the focal industry. Thus, the study estimates that a 10% increase in the labor productivity of an exporter for an industry leads to about a 20% expansion in export volumes within that industry compared to other industries for the exporter. This instrumented elasticity is weaker than Costinot et al.’s (2012) preferred estimate of 6.5 derived through producer price data for OECD countries in 1997, but it is quite similar to their 2.7 elasticity with labor productivity data that are most comparable to this study. The two analyses are also qualitatively similar in terms of their relationships to uninstrumented elasticities. This study does not find evidence of substantial adjustments 1. Channels for this technology transfer include communications among scientists and engineers (e.g., Saxenian 2002, Kerr 2008, Agrawal et al. 2011), trade flows (e.g., Rauch 2001, Rauch and Trindade 2002), and foreign direct investment (e.g., Kugler and Rapoport 2007, Foley and Kerr 2013). The supplement provides further references to the role of international labor mobility and other sources of heterogeneous technology frontiers (e.g., Eaton and Kortum 1999, Keller 2002). in the extensive margin of the group of countries to which the exporter trades. These results are robust to sample composition adjustments and variations on estimation techniques. Extensions quantify the extent to which heterogeneous technology transfer can be distinguished from a Rybczynski effect operating within manufacturing, evaluate differences in education levels or time in the United States for past migrants in instrument design, and test the robustness to controlling for direct ethnic patenting growth by industry in the United States. This study concludes that comparative advantages are an important determinant of trade; moreover, Ricardian differences are relevant for explaining changes in trade patterns over time. These panel exercises are closest in spirit to the industrial specialization work of Harrigan (1997) and the structural Ricardian model of Costinot et al. (2012). Other tests of the Ricardian model are MacDougall (1951, 1952), Stern (1962), Golub and Hsieh (2000), Chor (2010), Morrow (2010), Fieler (2011), Bombardini et al. (2012), Costinot and Donaldson (2012), Shikher (2012), Levchenko and Zhang (2014), Simonovska and Waugh (2014a,b), and Caliendo and Parro (2015). Recent related work on the industry dimension of trade includes Autor et al. (2013), Kovak (2013), and Hakobyan and McLaren (2016). Costinot and Rodriguez-Clare (2014) review empirical aspects and challenges of this literature. The comparative advantages of this work are in its substantial attention to non-OECD economies, the stricter panel assessment using heterogeneous technology diffusion, and the instruments built off of differential access to the US frontier. Work on migration-trade linkages dates back to Gould (1994), Head and Reis (1998), and Rauch and Trindade (2002), with Bo and Jacks (2012), di Giovanni et al. (2015), Bahar and Rapoport (2016), and Cohen et al. (2016) being recent contributions that provide references to the lengthy subsequent literature. This paper differs from these studies in its focus on technology transfer’s role for export promotion as an independent mechanism from migrant networks. In addition to contributing to the trade literature, the study documents for emerging economies an economic consequence of emigration to frontier economies like the United States. 2 2. Davis and Weinstein (2002) consider immigration to the United States, technology, and Ricardian- based trade. Their concern, however, is with the calculation of welfare consequences for US natives as a consequence of immigration due to shifts in trade patterns. I. ESTIMATING FRAMEWORK This section extends the basic estimating equation from Costinot et al. (2012) to a panel data setting. A simple application builds ethnic networks and heterogeneous technology diffusion into this theory. The boundaries of the framework and the statistical properties of the estimating equation are discussed. 3 Estimating Equation Costinot et al. (2012) develop a multi-country and multi-industry Ricardian model that has been widely studied and utilized in the trade literature. This framework builds off the model of Eaton and Kortum (2002) to articulate appropriate estimation of Ricardian advantages with industry-level data. The supplemental appendix shows how this model provides a microfoundation for studying Ricardian trade through an econometric specification of the form ln ( x ij k ) =δ ij + δ jk + θ ln( z ik ) + ε ijk , (1) where i indexes exporters, j indexes importers, and k indexes goods. Each good k has an infinite number of subvarieties that are being bought and sold with observed trade flows ij being an aggregation of the subvarieties. In the estimating equation, x k represents trade ik flows from exporter i to importer j for good k that adjust for country openness, and z represents observed labor productivity in exporter i for good k. As described in the supplemental appendix, the theory framework requires including fixed effects for bilateral trade routes ( δ ij ) and importer-industry fixed effects ( δ jk ) to account for unmodeled factors like consumer preferences, country sizes, and delivery costs. Finally, the estimated coefficient θ has a specific interpretation related to the Fréchet distribution that underlies this model and Eaton and Kortum (2002). Specifically, a low θ suggests a large scope for intraindustry comparative advantage, while a high θ (corresponding to 3. Dornbusch et al. (1977), Wilson (1980), Baxter (1992), Alvarez and Lucas (2007), Costinot (2009), and Costinot and Vogel (2015) provide further theoretical underpinnings for comparative advantage. large observed adjustments in exports with industry-level productivity shifts) suggests a limited scope for intraindustry comparative advantage. Estimates of θ in the trade literature have been derived with cross-sectional regressions using equation (1). This study seeks identification of the θ parameter within the Costinot et al. (2012) setting via first differencing and instrumental variables. 4 The first step is to extend equation (1) to include time t, ln ( x ijt k ) = δ ijt + δ jtk + θ ln( z k it ) + ε ijt k . (2) It is important to note that this extension is being applied to the fixed effect terms. Thus, the exporter-importer fixed effects in the cross-sectional format become exporter- importer-year fixed effects in a panel format. It is assumed that θ does not vary by period, although stacked versions of the Costinot et al. (2012) model could allow for this. The empirical work below estimates equation (2) for reference, but most of the specifications instead examine a first-differenced form, ∆ ln ( x ijt k )= δ ijt + δ jtk + θ∆ ln( z k it ) + ε ijt k , (3) where the fixed effects and error term are appropriately adjusted. The motivation for first differencing is stronger empirical isolation of the θ parameter. By themselves, exporter-importer-year and importer-industry-year fixed effects in equation (2) allow identification of the θ parameter in two ways: (i) k k it longitudinal changes in z it over time; and (ii) long-term differences in z across industries for the exporter. In a cross-sectional estimation of equation (1), it is not feasible to distinguish between these forms. This second effect persists when extending the equation (2) to a panel setting because the exporter-importer-year fixed effects δijt only account for the aggregate technology changes for exporters. First differencing best k it isolates the particular role of longitudinal changes in productivity z over time. 4. Daruich et al. (2016) estimate this framework encompasses about 20% of the variation in trade flows. Other studies seek to jointly model Ricardian advantages with other determinants of trade (e.g., Davis and Weinstein 2001, Morrow 2010). Whether estimating the θ parameter through both forms of variation is appropriate depends upon model assumptions, beliefs about unmeasured factors, and measurement error. It is helpful to illustrate by considering the exports of Germany in automobiles. The study examines trade over the 1980–1999 period. Throughout this period, Germany held strong technological advantages and labor productivity for manufacturing automobiles relative to the rest of the world. Over the course of the period, this productivity also changed in relative terms. If one can feasibly isolate these productivity variables, then having both forms of variation is an advantage. A second and related issue is that first differencing the data exacerbates the downward bias that measurement error causes for estimates of the θ parameter. There are plenty of reasons to suspect non-trivial measurement error in industry-level labor productivity estimates developed from the UNIDO database. On the other hand, removing time-invariant differences to identify the θ parameter can be an advantage. The basic identification constraint for the econometric analysis is that technology levels of exporters cannot be distinguished from other unobservable factors that also vary by exporter-industry or exporter-industry-year for the long-term technology levels and their longitudinal changes, respectively. The first is particularly worrisome given its general nature. First differencing is not foolproof against omitted factors, but it does require that the changes in these factors correlate with the changes in k it the focal productivity level in the exporters of z . This latter approach of panel estimation, while very common in microeconomic analyses, has yet to be extended to the Ricardian literature. 5 Beyond this discussion, a few other notes about the estimation of (3) are warranted. The dependent variable is bilateral manufacturing exports by exporter-importer-industry- year. The lack of trade for a large number of bilateral routes at the industry level creates 5. Estimations of the Costinot et al. (2012) model rely on fixed effects to handle delivery costs and other aspects of trade that are not due to the productivity of exporters. Thus, a cross-sectional estimation (1) requires unmodeled delivery costs be only comprised of a bilateral component and an importer-industry k component ( d ij = dij ·d k j ). A panel estimation (3) allows this proportionate structure to be extended to k dijt = dijt ·d k k jt ·d ij , where the third term represents the long-term delivery costs for the exporter to the importer by industry. econometric challenges with a log specification. These zero-valued exports are predicted by the model as an exporter is rarely the lowest-cost producer for all countries in an industry. This study approaches this problem by separately testing the intensive and extensive margins of trade. Most of the focus is on the intensive margin of trade expansion, where the dependent variable is the log growth in the value of bilateral exports ∆ ln ( x ijt k ) . The intensive margin of exports captures both quantities effects and price effects (e.g., Acemoglu and Ventura 2002, Hummels and Klenow 2005). In tests of extensive margin of trade expansion—that is, commencing exports to new import destinations—the dependent variable becomes a dichotomous indicator variable for whether measurable exports exist. Differences in the sample construction for these two tests are discussed when describing the trade dataset. Beyond the model’s background, the exporter-importer-year fixed effects perform several functions. They intuitively require that Germany’s technology expansion for auto manufacturing exceed its technology expansion for chemicals manufacturing if export growth is stronger in autos than chemicals. Thus, these fixed effects remove aggregate trade growth by exporter-importer pairs common across industries. These uniform expansions could descend from factors specific to one country of the pair (e.g., economic growth and business cycles, factor accumulations, terms of trade and price levels) or be specific to the bilateral trading pair (e.g., trade agreements, preferences 6). This framework is thus a powerful check against omitted variables biases, helping to isolate the Ricardian impetus for trade from relative factor scarcities and other determinants of trade. The fixed effects also control for the gravity covariates commonly used in empirical trade studies. National changes in factor endowments may still influence industries differentially due to the Rybczynski effect, which is explicitly tested for below. The importer-industry-year fixed effects control for tariffs imposed upon an industry in the importing country. More broadly, they also control for the aggregate growth in worldwide trade in each industry, relative price changes, and the potential for trade due to increasing returns to scale (e.g., Helpman and Krugman 1985, Antweiler and Trefler 2002). 6. Hunter and Markusen (1988) and Hunter (1991) find these stimulants account for up to 20% of world trade. More subtly, a key difference between multicountry Ricardian frameworks and the classic two-country model of Dornbusch et al. (1977) is worth emphasizing. This difference influences how the comparative static of increasing a single country-industry k it technology parameter z , ceteris paribus, is viewed. The multicountry theoretical k it framework allows for increases in z to reduce exports on some bilateral routes for the exporter-industry. This effect is due to general equilibrium pressures on input costs and extreme value distributions—while the productivity growth makes the exporter more competitive, the rising wage rates in the country may make it less competitive for a particular importer. This more nuanced pattern is different from the stark prediction of a two-country model where productivity growth in an industry for a country would never lead to declines in exports to the other country in that industry. The proper treatment effect for productivity growth is measured across all export destinations, as in the empirical work of this paper, and thus captures the general Ricardian pattern embedded in the model. This treatment effect is a net effect that may include reduction of exports on some routes. 7 Heterogeneous Technology Diffusion and Ricardian Trade While the Ricardian framework assigns a causal relationship of export growth to technology development, in practice the empirical estimation of specification (3) can be confounded by reverse causality or omitted variables operating by exporter-industry-year even after first differencing. Reverse causality may arise if engagement in exporting leads to greater technology adoption, perhaps through learning-by-doing or for compliance with an importer’s standards and regulations. An example of an exporter-industry-year omitted factor is a change in government policies to promote a specific industry, perhaps leading to large technology investments and the adoption of policies that favor the chosen industry’s exports relative to other manufacturing industries. This would lead to an upward bias in the estimated θ parameter. 8 7. Costinot et al. (2012) provide a more detailed discussion, including the extent to which the industry ordering of the two-country model is found in the relative ordering of exports for countries. 8. More specifically, the innovation in industrial policy support must be non-proportional across manufacturing industries. Long-term policies to support certain industries more than others are accounted Heterogeneous technology transfer from the United States provides an empirical foothold against these complications. Consider a leader-follower model where the technology state in exporter i and industry k is =  z k it tk ,US ·Υ ik ·Υ it ·M it z k . (4) tk ,US is the exogenously determined US technology frontier for each industry and year. z Two general shifters govern the extent to which foreign nations access this frontier. First, Υ ik models time-invariant differences in the access to or importance of US technologies to exporter i and industry k, potentially arising due to geographic separation (e.g., Keller 2002), heterogeneous production techniques (e.g., Davis and Weinstein 2001, Acemoglu and Zilibotti 2001), or similar factors. The shifter Υ it models longitudinal changes in the utilization of US technologies common to all industries within exporter i, such as changes due to declines in communication and transportation costs, greater general scientific or business integration, and so on. In what follows, both of these shifters could further be made specific to an exporter-importer pair. By themselves, these first three terms of model (4) describe the realities of technology diffusion but are not useful for identification when estimating specification (3). The tk ,US is captured by the importer-industry-year fixed effects, the technology frontier z bilateral Υ ik shifter is removed in the first differencing, and the longitudinal Υ it shifter is k captured in the exporter-industry-year fixed effects. The final term M it , however, describes differential access that the migrants to the United States from exporter i provide to the technologies used in industry k. This term models the recent empirical literature that finds that overseas diaspora and ethnic communities aid technology transfer from frontier countries to their home countries. If there is sufficient industry variation in this technology transfer, after removing the many fixed effects embedded into specification (3), then this transfer may provide an exogenous instrument to the exporter productivity for by the first differencing. Uniform changes in support across industries are also jointly accounted for by panel fixed effects. k it parameter z in a way that allows very powerful identification for the role of Ricardian advantages in trade. The design of this instrument combines spatial variation in historical settlement patterns in the United States of migrant groups from countries with spatial variation in where new technologies emerged over the period of the study. The instrument takes the form  Techck,,t A− S  k M it = ∑ M %i ,c ,1980 · k , A− S  , (5) c∈C  Techc ,1980    where c indexes US cities. M %i ,c ,1980 is the share of individuals tracing their ancestry to country i—defined in more detail below and including first-generation immigrants—that are located in city c in 1980. These shares sum to 100% across US cities. The bracketed fraction is a technology ratio defined for an industry k. The ratio measures for each city how much patenting grew in industry k relative to its initial level in 1980. The fraction exceeds one when a city’s level of invention for industry k grows from the base period, and it falls below one if the city’s invention for an industry weakens. The instrument thus interacts the spatial distribution across US cities of migrants from exporter i with the city-by-city degree to which technological development for industry k grew in locations. By summing across cities, equation (5) develops a total metric for exporter i and industry k that can be first differenced in a log format, ∆ ln( M it k ) , to instrument for ∆ ln( z k it ) in equation (3). A subtle but important point is that the instrument can only work in a first-differenced format (or equivalent panel data model with bilateral route fixed effects). This restriction is because the expression (5) does not have a meaningful cross-sectional level to it—for all countries and industries, the value of k k M it is equal to one in 1980 by definition. As such, M it cannot predict the cross-section k of trade in 1980. However, M it does provide insight about changes in technology opportunity sets over time that can be used for identification in estimations that consider changes in technology and trade over time. Three other points about the instrument’s design are important to bring out as they specifically relate to potential concerns about the instrument. First, the technology trend modeled in equation (5) is at the city-industry level, not at the city level. This is vital because the instrument interacts the US spatial ancestry distribution for a country with these city-industry patenting trends. As the estimations include exporter-importer-year fixed effects, any variable or instrument that would interact a city-level trend (e.g., population growth, housing prices, local public expenditures, etc.) with the city spatial distribution will be completely absorbed by the exporter-importer-year fixed effects (specifically, the exporter-year part of these fixed effects). In other words, general city- level factors like the total patenting of a location are anticipated to impact industries for a country in a proportionate way, and the estimations only use disproportionate variation over industries to identify the empirical effects. Second, one concern would be that migrants from exporter i select cities specifically to acquire technologies useful for their home country’s exports. This seems less worrisome perhaps for individual migrants, but it is quite plausible when contemplating a German automobile manufacturer opening a new facility in the United States (e.g., Alcacer and Chung 2007). The instrument seeks to rule out this concern by fixing the city distribution of migrants from exporter i at their city locations in 1980. This approach eliminates endogenous resorting, and the results below are also shown to be robust to focusing on second-generation and earlier migrants. Additional analyses also consider dropping industries for each country where the concerns could be most pronounced. A third concern is one of reverse causality. The United States relies extensively on immigrants for its science and engineering labor force, with first-generation immigrants accounting for about a quarter of the bachelor’s educated workforce and half of those with PhDs. Moreover, immigrants account for the majority of the recent growth in the US science and engineering workforce. The spatial patterns of new high-skilled immigrants frequently build upon ethnic enclaves and impact the innovation levels in those locations (e.g., Kerr and Lincoln 2010, Hunt and Gauthier-Loiselle 2010, Peri et al. 2015). Thus, a worry could be that the technology growth for cities in model (5) is endogenous. The concern would be that Germany is rapidly developing innovations and new technologies for the automobile industry, and this expansion is simultaneously leading to greater exports from Germany and the migration of German scientists that are patenting automobile technologies to the United States. This concern is addressed in several ways throughout this study, including sample decomposition exercises, lag structure tests, and similar exercises. The most straightforward safeguard, however, is already built into model (5). The patenting data, as described below, allow us to separate the probable ethnicities of inventors in the United States. By focusing on inventors of Anglo-Saxon ethnic heritage, one can remove much of this reverse causality concern. The Anglo-Saxon group accounts for about 70% of US inventors during the time period studied, and so this group reflects the bulk and direction of US technological development. Extensions will further consider settings where patent citation records suggest that the Anglo-Saxon inventors are mainly drawing on other Anglo-Saxon inventors in their research. 9 Addressing these concerns also provides the approach (5) with a conceptual advantage with respect to the fixed effect estimation strategy. The first differencing in specification (3) controls for the initial distributions M %i ,c ,1980 , and the importer- industry-year fixed effects δ jt k control for the technology growth ratio for industry k. This separation is not perfect due to the summation over cities, but it is closely mimicked. Thus, the identification in these estimations comes off these particular interactions. This provides a strong lever against concerns of omitted factors or reverse causality, and the well-measured US data can provide instruments that overcome the downward bias in coefficients due to measurement error. II. DATA PREPARATION 9. Very strong crowding-in or crowding-out of natives by immigrant scientists and engineers would create a bias in the Anglo-Saxon trend itself. Kerr and Lincoln (2010) find very limited evidence of either effect at the city level for the United States during this time period and for the time horizons considered here (i.e., first differencing over five-year periods). This section summarizes the key data employed in this study and their preparation, with the supplement providing further details. Table S1a describes the 88 exporting countries included, and table S1b provides similar statistics for the 26 industries, aggregating over countries. Labor Productivity Data k it Productivity measures z are taken from the Industrial Statistics Database of the United Nations Industrial Development Organization (UNIDO). The UNIDO data provide an unbalanced panel over countries, industries, and time periods, and the availability of these data are the key determinant of this study’s sample design. Estimations consider manufacturing industries at the three-digit level of the International Standard Industrial Classification system (ISIC3). Data construction starts by calculating the annual labor productivity in available industries and countries during the 1980–1999 period. These annual measures are then collapsed into the mean labor productivity level for each five- year period from 1980–1984 to 1995–1999. This aggregation into five-year time periods affords a more balanced panel by abstracting away from the occasional years when an otherwise reported country-industry is not observed. The higher aggregation is also computationally necessary below due to the tremendous number of fixed effects considered. These labor productivity measures are first differenced in log format for inclusion in equation (3). Thus, an exporter i and industry k is included if it is observed in the UNIDO database in two adjacent periods. Sample inclusion also requires that the country-industry be reported in two observations at least five years apart (e.g., to prevent an included observation only being present in 1989 and 1991). The main estimations consider the three change periods of 1980–1984 →1985–1989, 1985–1989 →1990–1994, and 1990– 1994 →1995–1999. Export Volumes ijt Bilateral exports x k are taken from the 1975–2000 World Trade Flows Database (WTF) developed by Feenstra et al. (2005). This rich data source documents product-level values of bilateral trade for most countries from 1980–1999. Similar to the development of the labor productivity variables, these product flows are aggregated into five-year periods from 1980–1984 to 1995–1999 and then first differenced in log format. Each productivity growth observation available with the UNIDO dataset is paired with industry-level bilateral export observations from that country. All exporting countries other than the United States are included. The majority of export volumes for bilateral routes are zero-valued, which creates challenges for the estimation of equation (3). It is also the case that the minimum threshold of trade that can be consistently measured across countries and industries is US $100,000 in the WTF database. While Feenstra et al. (2005) are able to incorporate smaller trading levels for some countries, these values are ignored to maintain a consistent threshold across observations. To accommodate these conditions, the empirical approach separately studies the extensive and intensive margins of export expansion. Mean export volumes are taken across exporter-importer-industry observations for five- year time periods. For the extensive margin, entry into exports along an exporter- importer-industry route is defined as exports greater than US $100,000. US Historical Settlement Patterns The first building block for the instrument is the historical settlement patterns of migrants from each country M %i ,c ,1980 . These data are taken from the 1980 Census of Populations, which is the earliest US census to collect the detailed ancestry of respondents (as distinguished from immigration status or place of birth). The detailed ancestry codes include 392 categories with positive responses, and this study maps these categories to the UNIDO records. Respondents are asked primary and secondary ancestries, but the classifications only focus on the primary field given the many missing values in the secondary field. There are multiple ancestry groups that map to the same country, but the mapping procedure limits each ancestry group to map to just one UNIDO country. Categories not linked to a specific UNIDO country are dropped (e.g., Western Europe not elsewhere classified, Ossetian). In total, 89% of the US population in 1980 is mapped. Metropolitan statistical areas, which will be referred to as cities for expositional ease, are identified using the 1% Metro Sample. This dataset is a 1-in-100 random sample of the US population in 1980 and is designed to provide accurate portraits of cities. The set C over which M %i ,c ,1980 is calculated includes 210 cities from the 1980 census files that are linked to the US patent data described next. The primary measures of M %i ,c ,1980 include all individuals regardless of age or education level to form M %i ,c ,1980 , only dropping those in group quarters (e.g., military barracks) or not living in an urban area. Extensions test variations on these themes. US Patenting Data The second building block for the instrument is the trend in patenting for each city Techck,,t A− S . These series are quantified through individual records of all patents granted by the United States Patent and Trademark Office (USPTO) from January 1975 to May 2009. Each patent record provides information about the invention (e.g., technology classification, citations of patents on which the current invention builds) and inventors submitting the application (e.g., name, city). USPTO patents must list at least one inventor, and multiple inventors are allowed. Approximately 7.8 million inventors are associated with 4.5 million granted patents during this period. The online supplement documents how the patent data are augmented in terms of city and industry definitions. Only patents with all inventors living in the United States at the time of their patent application are included, and multiple inventors are discounted so that each patent receives the same weight when measuring inventor populations. Concordances link USPTO technology classes to ISIC3 industries in which new inventions are manufactured or used. The main estimations focus on industry-of-use, affording a composite view of the technological opportunity developed for an industry. The probable ethnicities of inventors are estimated through the names listed on patents (e.g., Kerr 2007). This procedure exploits the fact that individuals with surnames Gupta or Desai are likely to be Indian, Wang or Ming are likely to be Chinese, and Martinez or Rodriguez are likely to be Hispanic. The name matching work exploits two commercial databases of ethnic first names and surnames, and the procedures have been extensively customized for the USPTO data. The match rate is 98% for US domestic inventors, and the process affords the distinction of nine ethnicities: Anglo-Saxon, Chinese, European, Hispanic, Indian, Japanese, Korean, Russian, and Vietnamese. Most of the estimations in this paper only use whether inventors are of Anglo-Saxon origin, as a means for reducing the potential of reverse causality as discussed above. The Anglo- Saxon share of US domestic patenting declines from 73% in 1980–1984 to 66% in 1995– 1999 (table S2). This group accounts for a majority of patents in each of the six major technology categories developed by Hall et al. (2001). As with the productivity and trade data, the patenting series are aggregated into five- year blocks by city and industry. These intervals start in 1975–1979 and extend through 1995–1999, and the series are normalized by the patenting level of each city-industry in 1980–1984. These series are then united with the spatial distribution of each country’s ancestry group using model (5) to form an aggregate for each country-industry, and the log growth rate is then calculated across these five-year intervals. The lag of this growth rate is used as the instrument for the productivity growth rate in an exporter-industry. That is, the estimated growth in technology flows from Brazil’s chemical industry during 1975–1979 →1980–1984 is used as the instrument for the growth in Brazil’s labor productivity in chemicals for the 1980–1984 →1985–1989 period. This lag structure follows the emphasis in Kerr (2008) on the strength of ethnic networks for technology diffusion during the first 3–6 years after a US invention is developed, and the comparison to contemporaneous flows is shown in robustness checks. The supplement provides additional notes on the instrument design and its connection to patent data. III. EMPIRICAL RESULTS This combined dataset is a unique laboratory for evaluating Ricardian technology differences in international trade. This section commences with ordinary least squares (OLS) estimations using the UNIDO and WTF data. The instrumental variable (IV) results are then presented. Base OLS Specifications Table 1 provides the basic OLS estimations. Column 1 presents the “between” estimates from specification (2) before first differencing the data; the dependent variable is the log mean nominal value of bilateral exports for the five-year period. These estimates identify the θ parameter through variation within bilateral trading routes and variation across industries of an exporter. This framework parallels most Ricardian empirical studies. Column 2 presents the “within” estimate from specification (3) that utilizes first differencing to isolate productivity and trade growth within exporter-importer-industry cells. Estimations in panel A weight bilateral routes by an interaction of total exporter and importer trade in the industry. For example, the weight given to Germany’s exports of automobiles to Nepal is the total export volume of Germany in the auto industry interacted with the total imports of Nepal in the auto industry, using averages for each component across the sample period. These weights focus attention on routes that are likely to be more important and give a sense of the overall treatment effect from Ricardian advantages. The weights, however, explicitly do not build upon the actual trade volume for a route to avoid an endogenous emphasis on where trade is occurring. Estimates in panel B are unweighted. This study reports results with both strategies to provide a range of estimates. Estimations cluster standard errors by exporter-industry. This reflects the repeated application of exporter-industry technology levels to each route and the serial correlations concerns of panel models. Other variants are reported below, too. Finally, the combination of 88 countries, 26 industries, and three time intervals creates an enormous number of exporter-importer-year and importer-industry-year fixed effects. The number of import destinations is in fact larger than the 88 exporters, as a UNIDO data match is not required for import destinations. With such a large dataset, it is computationally difficult to include exporter-importer-year and importer-industry-year fixed effects, especially when considering IV estimations. By necessity, manual demeaning is employed to remove the exporter-importer-year fixed effects, and this procedure is applied over the importer-industry-year fixed effects. The baseline estimates also use an aggregated version of the importer-industry-year fixed effects where the industry level used for the groups is at the two-digit level of the ISIC system rather than the three-digit level (reducing this dimension from 26 industries to eight higher-level industry groups). Robustness checks on these simplifications are reported below. Interestingly, the “between” and “within” elasticities estimated in panel A are both around 0.6 on the intensive margin. These coefficients suggest that a 10% growth in labor productivity for an exporter-industry is associated with a 6% growth in exports. The estimates in panel B are lower at 0.2–0.4, but they remain economically and statistically important. These elasticities are somewhat lower than the unit elasticity often found in this literature with OLS estimation techniques and cross-sectional data. There are many empirical reasons why this might be true, with greater measurement error for productivity estimates outside of OECD sources certainly being among them. An elasticity greater than or equal to one is also the baseline for the Ricardian theory presented earlier. The IV estimates reported below are greater than one and have a comparable level on some dimensions to those estimated with OECD countries. The next subsection continues with extensions for these OLS estimates to provide a foundation for the IV results. Extended OLS Results Table 2 provides robustness checks on the first-differenced estimates, which are the focus of the remainder of this study. The first column repeats the core results from column 2 of table 1. The next two columns show robustness to dropping Brazil and China. Brazil, of all included countries, displays the most outlier behavior with respect to its productivity growth rates, likely due to definitional changes, but Brazil’s exclusion does not affect the results. The results are also similar when excluding China, which experienced substantial growth during the sample period. It is generally worth noting that the 1980–1999 period predates the very rapid take-off of Chinese manufacturing exports after 2000 (Autor et al. 2013). Unreported tests consider other candidates like Mexico, Germany, and Japan, and these tests, too, find the results very stable to the sample composition, reflective in large part of the underlying exporter-importer-year fixed effects. Column 4 shows the results when excluding industry 383 (Machinery, electrical). The coefficient estimates are reduced in size by about 30% from column 1, but they remain quite strong and well-measured overall. The exclusion of industry 383 has the largest impact on the results of the 26 industries in the sample, which is why it is reported. This importance is not very surprising given the very rapid development of technology in this sector, its substantial diffusion around the world, and its associated trade. On this dimension, the industry-year portion of the importer-industry-year fixed effects play a very stabilizing role. Quantitatively similar results are also found when excluding the Tobacco and Petroleum sectors (314, 353, 354). Column 5 shows that winsorizing the sample at the 2%/98% level delivers similar results, indicative that outliers are not overly influencing the measured elasticities. It was earlier noted that computational demands require that the main estimations employ the ISIC2-level industry groups when preparing importer-industry-year fixed effects. Columns 6–8 test this choice in several ways. First, column 6 shows that the results hold when estimating the full model with ISIC2-based cells, so that the importer- industry-year fixed effects exactly match the cell construction. The weaker variation reduces the coefficient estimates by half, but the results remain statistically and economically important. Columns 7 and 8 alternatively estimate the model using the sample from Kerr (2008) that focuses on a subset of the UNIDO data in the 1985–1997 period. The Kerr (2008) sample is substantially smaller in size than the present one, and so there is greater flexibility with respect to these fixed effect choices. The choice of industry aggregation for the importer-industry-year fixed effects does not make a material difference in this sample. 10 Finally, column 9 shows the results with exporter-level clustering. The labor productivity and export development of industries within countries may be correlated with each other due to the presence of general-purpose technologies, learning-by-doing (e.g., Irwin and Klenow 1994), and similar factors, and Feenstra and Rose (2000) show how the export ranges of countries can change over time in systematic ways across industries. Clustering at the exporter level allows for greater covariance across industries in this regard, returning lower standard errors. Most papers in this literature use robust standard errors on cross-sectional data, which would translate most closely to bilateral- route clustering in a panel model. Unreported estimates consider bilateral-route clustering and alternatively bootstrapped standard errors, and these standard errors are smaller than those reported in column 9. 10. This extra check also has the advantage of linking the two studies closer together since the Kerr (2008) paper focuses extensively on productivity growth due to technology transfer. Stability to the somewhat different data preparation steps in Kerr (2008) is comforting. Further extensions are contained in tables S3 and S4. Interacting our core regressor with the GDP/capita level of the exporter suggests that the OLS link between productivity and exports is mainly through lower-income countries, suggestive of higher trade due to varieties among developed economies. It is similar to the conclusion of Fieler (2011) that trade among advanced economies links to product differentiation and variety (low θ ), while trade among emerging economies links more closely to fundamental productivity levels (higher θ). By contrast, there is limited heterogeneity by country size or geographic distances, excepting the fact that the growth in exports is not simply happening to bordering countries. Additional tests further confirm that the observed role for technology within manufacturing is not due to specialized factor accumulations and a Rybczynski effect. Under the Rybczynski effect, the accumulation of skilled workers in country i shifts country i’s specialization toward manufacturing industries that employ skilled labor more intensively than other factors. When incorporating country-specific time trends for subgroups of manufacturing industries according to their capital-labor ratio, mean wage rate, and non-production worker share as evident in the United States, technology’s importance is confirmed. Finally, there is limited adjustment on the extensive margin of trade routes compared to the intensive margin adjustments. Base IV Results Table 3 presents the core IV results. The first column reports the first-stage estimates of how ∆ ln( M it k ) predicts ∆ ln( z k it ) . The first-stage elasticity in panel A is 0.6, suggesting a 10% increase in the technology flow metric from the United States predicts a 6% increase in labor productivity abroad at the exporter-industry level. The unweighted estimates in panel B suggest a smaller 3% increase. While the second elasticity is lower, the instrument generally performs better in the unweighted specifications due to its more precise measurement. The F statistics in panels A and B are 4.7 and 11.6, respectively. The sample weights in panel A place greater emphasis on larger and more advanced countries that have large export volumes (e.g., Germany, Japan). While this framework finds a substantial response, the weighted dependency of this group on heterogeneous technology transfer from the United States is noisier than in the unweighted estimations that emphasize more developing and emerging countries. The second column presents the reduced-form estimates where ∆ ln( M it k ) predicts ∆ ln( x ijt k ) using a format similar to equation (3). In both panels, there is substantial reduced-form link of technology flows to export volumes. The interquartile range in the reduced form, conditional on the fixed effects, can account for around 6% of the interquartile range of export growth using a 0.8 coefficient estimate that sits in between panels A and B. The third column provides the second-stage estimates from equation (3) having used ∆ ln( M it k ) to predict ∆ ln( z k it ) . In panel A’s estimation, the weighted elasticity is 1.6, suggesting a 16% increase in export volumes for every 10% increase in labor productivity. In panel B’s unweighted estimation, the 10% increase in labor productivity is linked to a 24% increase in export volumes. The second-stage elasticity in panel B is larger than in panel A, as the IV estimates provide the reduced-form scaled up by the first-stage effects. Thus, even though the unweighted reduced-form estimate in column 2 is smaller than the weighted reduced-form estimate, this ordering reverses once scaled-up by the first stages. This study does not overly favor one set of estimates. The weighted and unweighted approaches both have merits and liabilities. Instead, the conclusion from this work is that the instrumented elasticity is in the neighborhood of two. While it is impossible to differentiate among the various reasons as to why the IV estimates are larger than the OLS estimates, a very likely candidate is that OLS suffers from a substantial downward bias due to measurement error in the labor productivity estimates, especially with the substantial differencing embedded in equation (3). While it is likely that omitted factors or reverse causality influenced the OLS estimations as well, these appear to have been second-order to the measurement issues. 11 11. Costinot et al. (2012) adjust export volumes for trade openness using the import penetration ratio for a country-industry (to link observed productivity to “fundamental productivity”). The estimates are very similar when undertaking this approach, being 1.163 (0.457) and 2.513 (1.138) for weighted and unweighted specifications, respectively. The unadjusted and adjusted first differences have a 0.93 correlation. This approach is not adopted for the main estimations due to worries about mismeasurement in the import penetration ratio when combining UNIDO and WTF data. This instrumented θ elasticity is at the lower end of the estimates provided in the literature. Costinot et al. (2012) is the closest comparison given their use of industry-level regressions of productivity data and trade. Using a cross-sectional analysis of producer price data for OECD countries in 1997, they derive their preferred estimate of 6.5, which is substantially larger than this study’s estimate of about two. On the other hand, Costinot et al. (2012) derive a quite similar elasticity of 2.7 when they consider labor productivity metrics, the metric considered here. They too have IV estimates that are considerably larger than OLS estimates. More broadly, Eaton and Kortum (2002) provide larger initial estimates of the θ elasticity, with a preferred estimate in the range of eight. Simonovska and Waugh (2014a) revisit these results with a new estimator and come to a preferred estimate of about four. Overall, this study’s estimates are again lower than this connected work. The robustness checks described below find θ estimates that continue in this ballpark, never exceeding four. Thus, this empirical approach consistently derives θ estimates that are among the lowest in the literature, with some part of this difference due to methodology but an important part being substantive in interpretation. It is important to identify the dual meaning of the higher IV results compared to OLS with respect to the θ parameter. In the Ricardian model, a higher θ parameter corresponds to a reduced scope for intraindustry trade due to comparative advantages across varieties. IV estimations thus suggest that OLS specifications overestimate the scope for intraindustry trade because they understate the link between country-industry productivity improvements and their associated export volumes. Both impetuses can be connected to Ricardian theories of comparative advantage for trade, but the role of the structural θ parameter needs to be carefully delineated. This partitioning can also have important consequences for views of development and export success. Compared to OLS, the IV results shift more emphasis toward fundamental country-industry productivity improvements rather than intraindustry varieties; yet, on the whole, the overall work in this paper with emerging economies provides more support for intraindustry varieties than typically found in the Ricardian literature that has focused mostly on OECD trade flows (as evidenced by the lower θ estimates compared to prior studies). These estimates are significant in terms of their potential economic importance and explanatory power. Using an elasticity of two, the interquartile range of country-level labor productivity growth, conditional on fixed effects, can explain up to 35% and 39% of the interquartile range in conditional export growth levels using unweighted and weighted specifications, respectively. Extended IV Results The supplement provides many robustness checks on these IV estimations in tables S5- S8b. The results are robust to the various specification checks conducted in table 2. Moreover, dynamic estimations find lagged productivity growth estimators consistently stronger than contemporaneous productivity growth estimators, providing comfort in the estimation design and the proposed causal direction of the results. The IV is further analyzed when: (i) considering variations on the city-industry technology trend terms and k cross-sectional distributions used to weight US cities for the development of M it ; (ii) sample composition exercises that aggressively test for data quality and reverse causality concerns; and (iii) direct inclusion of ethnic patenting in the United States as a control. The supplement also contains an extended discussion of the identification achieved in this study and its limitations. In the end, the paper is able to make substantial progress toward causal identification in a Ricardian model, beginning with the panel estimation approach and extending through many IV approaches. This effort remains incomplete, however, and it is hoped that future work both identifies natural experiment settings to test these arguments and also identifies other forms/impetuses for heterogeneous technology transfer that can provide identification in a setting that focuses on large country-industry samples like this one. IV. CONCLUSIONS While the principle of Ricardian technology differences as a source of trade is well established in the theory of international economics, empirical evaluations of its importance are relatively rare due to the difficulty of quantifying and isolating technology differences. This study exploits heterogeneous technology diffusion from the United States through ethnic migrant networks to make additional headway. Estimations find bilateral manufacturing exports respond positively to growth in observable measures of comparative advantages. Ricardian technology differences are an important determinant of trade in longitudinal changes in addition to their cross-sectional role discussed earlier. Leamer and Levinsohn (1995) argue that trade models should be taken with a grain of salt and applied in contexts for which they are appropriate. This is certainly true when interpreting these results. The estimating frameworks have specifically sought to remove trade resulting from factor endowments, increasing returns, consumer preferences, and so on, rather than test against them. Moreover, manufacturing exports are likely more sensitive to patentable technology improvements than the average sector, and the empirical reach of the constructed dataset to include emerging economies like China and India heightens this sensitivity. Further research is needed to generalize technology’s role to a broader set of industrial sectors and environments. Beyond quantifying the link between technology and trade for manufacturing, this paper also serves as input into research regarding the benefits and costs of emigration to the United States for the migrants’ home countries (i.e., the “brain drain” or “brain gain” debate). While focusing on the Ricardian model and its parameters, the paper establishes that the technology transfers from overseas migrants are strong enough to meaningfully promote exports. Care should be taken to not overly interpret these findings as strong evidence of a big gain from migration. The paper does not seek to establish a clear counterfactual in the context of immigration from the source countries’ point of view (e.g., Agrawal et al. 2011). As such, the positive export elasticities due to US heterogeneous technology diffusion do not constitute welfare statements relative to other scenarios. Future research needs to examine these welfare implications further. REFERENCES Acemoglu, D., and J. Ventura. 2002. “The World Income Distribution.” Quarterly Journal of Economics 117 (2): 659–94. Acemoglu, D., and F. Zilibotti. 2001. “Productivity Differences.” Quarterly Journal of Economics 116 (2): 563–606. Agrawal, A., D. Kapur, J. McHale, and A. Oettl. 2011. “Brain Drain or Brain Bank? The Impact of Skilled Emigration on Poor-Country Innovation.” Journal of Urban Economics 69 (1): 43–55. Alcacer, J., and W. Chung. 2007. “Location Strategies and Knowledge Spillovers.” Management Science 53 (5): 760–76. Alvarez, F., and R. Lucas. 2007. “General Equilibrium Analysis of the Eaton-Kortum Model of International Trade.” Journal of Monetary Economics 54: 1726–68. Antweiler, W., and D. Trefler. 2002. “Increasing Returns to Scale and All That: A View From Trade.” American Economic Review 92 (1): 93–119. Autor, D., D. Dorn, and G. Hanson. 2013. “The China Syndrome: Local Labor Market Effects of Import Competition in the United States.” American Economic Review 103 (6): 2121–68. Bahar, D., and H. Rapoport. 2016. “Migration, Knowledge Diffusion, and the Comparative Advantages of Nations.” Economic Journal forthcoming. Baxter, M. 1992. “Fiscal Policy, Specialization, and Trade in the Two-Sector Model: The Return of Ricardo?” Journal of Political Economy 100 (4): 713–44. Bo, C., and D. Jacks. 2012. “Trade, Variety and Immigration.” Economic Letters 117 (1): 243–6. Bombardini, M., C. Kurz, and P. Morrow. 2012. “Ricardian Trade and the Impact of Domestic Competition on Export Performance.” Canadian Journal of Economics 45 (2): 585–612. Caliendo, L., and F. Parro. 2015. “Estimates of the Trade and Welfare Effects of NAFTA.” Review of Economic Studies 82 (1): 1–44. Chor, D. 2010. “Unpacking Sources of Comparative Advantage: A Quantitative Approach.” Journal of International Economics 82 (2): 152–67. Cohen, L., U. Gurun, and C. Malloy. 2016. “Resident Networks and Firm Trade.” Journal of Finance 6: 63–91. Costinot, A. 2009. “An Elementary Theory of Comparative Advantage.” Econometrica 77 (4): 1165–92. Costinot, A., and A. Rodriguez-Clare. 2014. “Trade with Numbers: Quantifying the Consequences of Globalization.” In Handbook of International Economics Volume 4. Amsterdam: Elsevier. Costinot, A., and D. Donaldson. 2012. “Ricardo’s Theory of Comparative Advantage: Old Idea, New Evidence.” American Economic Review Papers and Proceedings 102 (3): 453–8. Costinot, A., D. Donaldson, and I. Komunjer. 2012. “What Goods Do Countries Trade? New Ricardian Predictions.” Review of Economic Studies 79 (2): 581–608. Costinot, A., and J. Vogel. 2015. “Beyond Ricardo: Assignment Models in International Trade.” Annual Review of Economics 7: 31–62. Daruich, D., W. Easterly, and A. Reschef. 2016. “The Surprising Instability of Export Specializations.” NBER Working Paper 22869. Cambridge, Mass.: National Bureau of Economic Research. Davis, D., and D. Weinstein. 2001. “An Account of Global Factor Trade.” American Economic Review 91 (5): 1423–53. ———. 2002. “Technological Superiority and the Losses from Migration.” NBER Working Paper 8971. Cambridge, Mass.: National Bureau of Economic Research. di Giovanni, J., A. Levchenko, and F. Ortega. 2015. “A Global View of Cross-Border Migration.” Journal of the European Economic Association 31 (1): 168–202. Dornbusch, R., S. Fischer, and P. Samuelson. 1977. “Comparative Advantage, Trade and Payments in a Ricardian Model with a Continuum of Goods.” American Economic Review 67 (5): 823–39. Eaton, J., and S. Kortum. 1999. “International Technology Diffusion: Theory and Measurement.” International Economic Review 40 (3): 537–70. Eaton, J., and S. Kortum. 2002. “Technology, Geography, and Trade.” Econometrica 70 (5): 1741–79. Feenstra, R., R. Lipsey, H. Deng, A. Ma, and H. Mo. 2005. “World Trade Flows: 1962– 2000.” NBER Working Paper 11040. Cambridge, Mass.: National Bureau of Economic Research. Feenstra, R., and A. Rose. 2000. “Putting Things in Order: Trade Dynamics and Product Cycles.” Review of Economics and Statistics 82 (3): 369–82. Fieler, A.C. 2011. “Nonhomotheticity and Bilateral Export Trade: Evidence and a Quantitative Explanation.” Econometrica 79 (4): 1069–101. Foley, C.F., and W. Kerr. 2013. “Ethnic Innovation and U.S. Multinational Firm Activity.” Management Science 59 (7): 1529–44. Golub, S., and C. Hsieh. 2000. “Classical Ricardian Theory of Comparative Advantage Revisited.” Review of International Economics 8 (2): 221–34. Gould, D. 1994. “Immigrant Links to the Home Country: Empirical Implications for US Bilateral Trade Flows.” Review of Economics and Statistics 76 (3): 500–18. Hakobyan, S., and J. McLaren. 2016. “Looking for Local Labor Market Effects of NAFTA.” Review of Economics and Statistics 98 (4): 728–41. Hall, B., A. Jaffe, and M. Trajtenberg. 2001. “The NBER Patent Citation Data File: Lessons, Insights and Methodological Tools.” NBER Working Paper 8498. Cambridge, Mass.: National Bureau of Economic Research. Harrigan, J. 1997. “Technology, Factor Supplies, and International Specialization: Estimating the Neoclassical Model.” American Economic Review 87 (4): 475–94. Head, K., and J. Reis. 1998. “Immigration and Trade Creation: Econometric Evidence from Canada.” Canadian Journal of Economics 31 (1): 47–62. Helpman, E., and P. Krugman. 1985. Market Structure and Foreign Trade. Cambridge, Mass.: MIT Press. Hummels, D., and P. Klenow. 2005. “The Variety and Quality of a Nation’s Exports.” American Economic Review 95 (3): 704–23. Hunt, J., and M. Gauthier-Loiselle. 2010. “How Much Does Immigration Boost Innovation?” American Economic Journal: Macroeconomics 2 (2): 31–56. Hunter, L. 1991. “The Contribution of Non-Homothetic Preferences to Trade.” Journal of International Economics 30 (3): 345–58. Hunter, L., and J. Markusen. 1988. “Per Capita Income as a Determinant of Trade.” in R. Feenstra, ed., Empirical Methods for International Trade. Cambridge, Mass.: MIT Press. Irwin, D., and P. Klenow. 1994. “Learning by Doing Spillovers in the Semiconductor Industry.” Journal of Political Economy 102: 1200–27. Keller, W. 2002. “Geographic Localization of International Technology Diffusion.” American Economic Review 92 (1): 120–42. Kerr, W. 2008. “Ethnic Scientific Communities and International Technology Diffusion.” Review of Economics and Statistics 90 (3): 518–37. ———. 2007. “The Ethnic Composition of US Inventors.” HBS Working Paper 08-006. Boston, Mass.: Harvard Business School. ———. 2005. “The Role of Immigrant Scientists and Entrepreneurs in International Technology Transfer.” Ph.D. dissertation, MIT Cambridge, Mass. Kerr, W., and W. Lincoln. 2010. “The Supply Side of Innovation: H-1B Visa Reforms and US Ethnic Invention.” Journal of Labor Economics 28 (3): 473–508. Kovak, B. 2013. “Regional Effects of Trade Reform: What is the Correct Measure of Liberalization?” American Economic Review 103 (5): 1960–76. Kugler, M., and H. Rapoport. 2007. “International Labor and Capital Flows: Complements or Substitutes?” Economics Letters 92 (2): 155–62. Leamer, E., and J. Levinsohn. 1995. “International Trade Theory: The Evidence.” Handbook of International Economics 3: 1339–94. Levchenko, A., and J. Zhang. 2014. “Ricardian Productivity Differences and the Gains from Trade.” European Economic Review 65: 45–65. MacDougall, G.D.A. 1951. “British and American Exports: A Study Suggested by the Theory of Comparative Costs. Part I.” Economic Journal 61 (244): 697–724. ———. 1952. “British and American Exports: A Study Suggested by the Theory of Comparative Costs. Part II.” Economic Journal 62 (247): 487–521. Morrow, P. 2010. “Ricardian-Heckscher-Ohlin Comparative Advantage: Theory and Evidence.” Journal of International Economics 82 (2): 137–51. Peri, G., K. Shih, and C. Sparber. 2015. “STEM Workers, H1B Visas and Productivity in US Cities.” Journal of Labor Economics 33 (S1): S225–55. Rauch, J. 2001. “Business and Social Networks in International Trade.” Journal of Economic Literature 39 (4): 1177–1203. Rauch, J., and V. Trindade. 2002. “Ethnic Chinese Networks in International Trade.” Review of Economics and Statistics 84 (1): 116–30. Ricardo, D. 1817. On the Principles of Political Economy and Taxation. London, UK: John Murray. Saxenian, A., Y. Motoyama, and X. Quan. 2002. Local and Global Networks of Immigrant Professionals in Silicon Valley. San Francisco, CA: Public Policy Institute of California. Saxenian, A. 2006. The New Argonauts. Cambridge, Mass.: Harvard University Press. Shikher, S. 2012. “Putting Industries into the Eaton–Kortum Model.” Journal of International Trade & Economic Development 21 (6): 807–37. Simonovska, I., and M. Waugh. 2014a. “The Elasticities of Trade: Estimates and Evidence.” Journal of International Economics 92 (1): 34–50. ———. 2014b. “Trade Models, Trade Elasticities, and the Gains from Trade.” NBER Working Paper 20495. Cambridge, Mass.: National Bureau of Economic Research. Stern, R. 1962. “British and American Productivity and Comparative Costs in International Trade.” Oxford Economic Papers 14 (3): 275–96. Wilson, C. 1980. “On the General Structure of Ricardian Models with a Continuum of Goods: Applications to Growth, Tariff Theory, and Technical Change.” Econometrica 48 (7): 1675–1702. Table 1: OLS estimations of labor productivity and exports Between estimation FD estimation (1) (2) Panel A: Weighting bilateral routes by the interaction of exporter and importer trade in industry (summed across all bilateral routes) DV: Log bilateral exports DV: Δ Log bilateral exports Log country-industry labor 0.640*** productivity (0.242) Δ Log country-industry labor 0.573*** productivity (0.185) Observations 149,547 103,839 Importer-Industry-Yr FE Yes Yes Exporter-Importer-Yr FE Yes Yes Panel B: Excluding sample weights DV: Log bilateral exports DV: Δ Log bilateral exports Log country-industry labor 0.361*** productivity (0.091) Δ Log country-industry labor 0.210*** productivity (0.041) Observations 149,547 103,839 Importer-Industry-Yr FE Yes Yes Exporter-Importer-Yr FE Yes Yes Source: Author's analysis based on data described in text. Notes: Panel estimations consider manufacturing exports taken from the WTF database. Data are organized by exporter-importer-industry-year. Industries are defined at the three-digit level of the ISIC Revision 2 system. Annual data are collapsed into five-year groupings beginning with 1980-1984 and extending to 1995-1999. The dependent variable in Column 1 is the log mean nominal value (US$) of bilateral exports for the five years; the dependent variable in Column 2 is the change in log exports from the prior period. The intensive margin sample is restricted to exporter-importer-industry groupings with exports exceeding $100k in every year. The $100k threshold is chosen due to WTF data collection procedures discussed in the text. Labor productivity from the UNIDO database measures comparative advantages. Column 1 estimates Ricardian elasticities using both within-panel variation and variation between industries of a country. Column 2 estimates Ricardian elasticities using only variation within panels. Estimations in Panel A weight bilateral routes by the interaction of total exporter and importer trade in industry; estimations in Panel B are unweighted. Estimations cluster standard errors by exporter-industry. Importer-Industry-Yr FE are defined at the two-digit level of the ISIC system. *, **, *** denote statistical significance at the 10%, 5%, and 1% levels, respectively. Table 2: Robustness checks on OLS specifications in Table 1 Kerr (2008) Kerr (2008) Base Using a Using ISIC sample sample Using estimation Excluding Excluding Excluding 2%/98% 2-digit level using Imp- using Imp- exporter- (Column 2, exports exports electrical winsorized industry ISIC2-Year ISIC3-Year level Table 1) from Brazil from China machinery sample groups fixed fixed clustering (1) (2) (3) (4) (5) (6) (7) (8) (9) The dependent variable is Δ log bilateral exports on the intensive margin by exporter-importer-industry Panel A: Weighting bilateral routes by the interaction of exporter and importer trade in industry Δ Log country-industry labor 0.573*** 0.573*** 0.472*** 0.390*** 0.493*** 0.266** 0.287** 0.281** 0.573*** productivity (0.185) (0.185) (0.097) (0.121) (0.133) (0.112) (0.113) (0.138) (0.086) Observations 103,839 103,010 101,221 97,326 103,839 51,483 23,345 23,345 103,839 Importer-Industry-Yr FE Yes Yes Yes Yes Yes Yes Yes Yes Yes Exporter-Importer-Yr FE Yes Yes Yes Yes Yes Yes Yes Yes Yes Panel B: Excluding sample weights Δ Log country-industry labor 0.210*** 0.210*** 0.221*** 0.154*** 0.264*** 0.097*** 0.248*** 0.184*** 0.210*** productivity (0.041) (0.041) (0.041) (0.039) (0.043) (0.037) (0.066) (0.069) (0.048) Observations 103,839 103,010 101,221 97,326 103,839 51,483 23,345 23,345 103,839 Importer-Industry-Yr FE Yes Yes Yes Yes Yes Yes Yes Yes Yes Exporter-Importer-Yr FE Yes Yes Yes Yes Yes Yes Yes Yes Yes Source: Author's analysis based on data described in text. Notes: See Table 1. Table 3: IV estimations of labor productivity and exports First-stage estimation Reduced-form estimation IV estimation (1) (2) (3) Panel A: Weighting bilateral routes by the interaction of exporter and importer trade in industry DV: Δ Log country- DV: Δ Log bilateral DV: Δ Log bilateral industry labor productivity exports exports Δ Log estimator for technology 0.589** 0.938*** flows from the United States (0.272) (0.298) Δ Log country-industry labor 1.592** productivity (0.637) Observations 103,839 103,839 103,839 Importer-Industry-Yr FE Yes Yes Yes Exporter-Importer-Yr FE Yes Yes Yes Panel B: Excluding sample weights DV: Δ Log country- DV: Δ Log bilateral DV: Δ Log bilateral industry labor productivity exports exports Δ Log estimator for technology 0.267*** 0.648*** flows from the United States (0.078) (0.112) Δ Log country-industry labor 2.429*** productivity (0.791) Observations 103,839 103,839 103,839 Importer-Industry-Yr FE Yes Yes Yes Exporter-Importer-Yr FE Yes Yes Yes Source: Author's analysis based on data described in text. Notes: See Table 1. The instrument combines panel variation on the development of new technologies across US cities during the 1975-2000 period with historical settlement patterns for migrants and their ancestors from countries that are recorded in the 1980 Census of Populations. The F statistics in Panels A and B are 4.7 and 11.6, respectively. S1. Online Supplement for Heterogeneous Technology Di¤usion and Ricardian Trade Patterns I Additional Literature Recent research quanti…es the role of international labor mobility in ethnic technology exchanges (e.g., Saxenian 2006, Kapur and McHale 2005, Obukhova 2008, 2009, Nanda and Khanna 2010). Related unpublished work includes Papageorgiou and Spilimbergo (2008) and Kugler and Rapoport (2012). Other sources of heterogeneous technology frontiers are geographic dis- tances to major R&D nations (e.g., Keller 2002), the innovative e¤orts of trading partners (e.g., Grossman and Helpman 1991, Coe and Helpman 1995, Coe et al. 1997), or international patent- ing decisions (e.g., Eaton and Kortum 1999). Keller (2004) reviews the technology transfer literature. II Costinot et al. (2012) Model This theoretical framework derives from Costinot et al. (2012). A world economy is comprised of I countries and K industries or goods. Labor is the sole factor of production and there are constant returns to scale in the production of each good. Labor is perfectly mobile across industries and immobile across countries. Li and wi are the number of workers and the wage rate in country i, respectively. Consumers consume their full wages in each period. Accordingly, time subscripts are omitted until the estimating equation is introduced. Countries are free to produce or trade all goods. Each good k 2 K has an in…nite number of varieties indexed by !2 [1; : : : ; +1]. zik (! ) represents the number of units of the ! th variety of good k that can be produced with one unit of labor in country i. Following Eaton and Kortum (2002), zik (! ) is a random variable drawn independently for each triplet (i; k; ! ) from a Fréchet distribution, Fik (z ) = exp[ (z=zik ) ], for all z 0, (S1.1) where zik > 0 and > 1. Thus, technological di¤erences across countries and industries depend on two parameters, zik and . The …rst parameter zik captures the fundamental productivity of country i in industry k , which a¤ects the productivity of all producers (e.g., institutions, climate). For each industry, the cross-country variation of this zik parameter governs the cross- country variation in relative labor productivity that sits at the core of the standard Ricardian 1 model. A larger zik raises the absolute advantage for trade for exporter i in industry k . The second parameter models the intra-industry heterogeneity that exists due to the scope for idiosyncratic di¤erences in technological know-how across varieties. This variation is the same in all countries and industries, and parameterizes the impact of changes in fundamental productivity levels zik on aggregate trade ‡ows. A larger implies a tighter distribution that limits the scope for comparative advantage across nations. Trade frictions have iceberg costs such that for each unit of good k shipped from exporter i to importer j , only 1=dk ij 1 units arrive, with dk k ii = 1 and dil dk k ij djl for any third country l to rule out cross-country arbitrage opportunities. Perfect competition in markets and constant returns to scale in production imply that the price pk j (! ) paid by buyers of variety ! of good k in any country j is pk k j (! ) = min[cij (! )], (S1.2) i2 I where ck k k ij (! ) = (dij wi )=zi (! ) is the cost of producing and delivering one unit of this variety from country i to country j . For each variety ! of good k , buyers in country j select the best s e¢ ciency for good j lowers the price available from around the world. An increase in country i’ price it must charge. Representative consumers in each country have a two-tier utility function. The upper tier is k Cobb-Douglas, and the preference parameter j measures the share of expenditure on varieties from industry k in country j . The lower tier is constant elasticity of substitution (CES), and k j is the elasticity of substitution between varieties. Accordingly, expenditures are such that in any importer j , total expenditure on variety ! of good k is k 1 k xk k j (! ) = [pj (! )=pj ] j k j wj Lj , (S1.3) hP k i1=(1 k) j k k 0 1 where 0 j 1, j < 1 + , and pk j !0 2 pk j (! ) j . The restriction k j < 1 + is a technical assumption that guarantees the existence of a well-de…ned CES price index pk j . The QK k j k consumer price index in country j is pj k=1 (pj ) : P The value of total exports from exporter i to importer j in industry k is xk ij k 0 ! 0 2 k xj (! ), ij k where denotes the set of varieties exported. The share of exports in importer j and industry ij PI k from country i is k ij xk ij = k i0 =1 xi0 j . With this model structure, the bilateral exports from exporter i to importer j in industry k is (wi dk k ij =zi ) xk ij = PI k k k j wj Lj ; (S1.4) i0 =1 (wi0 di0 j =zi0 ) 2 which has an intuitive interpretation that closely connects to a similar expression in Eaton and Kortum (2002). The righthand terms express the overall economic size of the importer j and its preferences over goods. The lefthand fraction describes the extent to which the exporter is the lowest cost producer of the good, taking into account delivery costs, production costs, technology levels, and the underlying heterogeneity in varieties for countries. Under assumptions of balanced trade, the relatives wages around the world can further be determined. Costinot et al. (2012) show that equation (S1.4) provides the foundation for estimating an k ~k econometric equation of the form shown in speci…cation (1) of main text: ln x ij = ij + j + zik ) + "k ln(~ ~k ij , where x ij represents “corrected” trade ‡ows that adjust for country openness. ~ik represents observed productivity, given that not every country produces every good Similarly, z k as it can import goods from other countries. ij and j are vectors of exporter-importer and importer-industry …xed e¤ects. Comparing these equations shows the basic function of these …xed e¤ects. The importer-industry …xed e¤ects control for the righthand terms about the importer and its preferences over goods. Importer-industry …xed e¤ects also account for the denominator of the lefthand fraction, given that it is a worldwide aggregate for an industry. The s terms, which emphasize cost levels in exporter-importer …xed e¤ects capture the numerator’ the exporter and distances between the two countries. Under the assumption that the delivery cost term dk ij in the numerator can be expressed in proportionate terms over these two vectors of …xed e¤ects (e.g., dk k ij = dij dj ), speci…cation (1) of main text provides an unbiased estimate of the parameter. An alternative assumption is that the residual di¤erences in delivery costs after controlling for these vectors of …xed e¤ects are uncorrelated with the focal productivity ~ik . level z III Extended Data Development Labor Productivity Data The UNIDO collects industry-level manufacturing statistics for The International Yearbook of Industrial Statistics and specialized publications on topics like development and competition. Researchers at the UNIDO supplement the data resources of the OECD with national records s stated objective for non-OECD members, creating a unique global resource. The UNIDO’ is the compilation of internationally comparable and internally consistent series (e.g., variable de…nitions, accounting units, collection procedures). Table S1a describes the 88 exporting countries included. Column 2 provides a count of the 3 number of periods the country is included in after the …rst di¤erence is taken, with a maximum of three changes. Column 3 documents the count of bilateral route observations included at the exporter-importer-industry level for the intensive-margin estimations. The total observation count is 103,839 intensive-margin changes for an exporter-importer-industry. Countries di¤er in their observation counts, even if observed for the same number of periods, due to variations in their industry-level reporting in the UNIDO database and minimum requirements for export volumes discussed below. Column 4 documents the average manufacturing productivity levels for countries, expressed in US dollars. While direct comparisons across countries are limited with an unbalanced panel, productivity di¤erences between industrialized countries and developing nations are clearly evi- dent. Oil-producing countries (e.g., Kuwait, Norway) have the highest average labor productiv- ity levels, with Ireland, Japan and Singapore among the highest when excluding oil producers. Afghanistan, Bangladesh, and Myanmar are among the lowest levels recorded. A small num- ber of country-industry observations with under ten employees or very problematic data are excluded. Column 5 documents the mean growth rate in labor productivity for each country over its observations. These growth rates are …ve-year di¤erences, with outliers winsorized at their 2% and 98% levels for reporting. Hong Kong, Myanmar, Peru, and Syria have the highest growth rates, while the Dominican Republic, Tanzania, and Romania show the sharpest declines. Table S1b provides similar statistics for the 26 industries, aggregating over countries. In- dustry 353 (Petroleum re…neries) has the highest average labor productivity, while industry 322 (Wearing apparel, except footwear) has the lowest. Productivity growth is strongest in industries 382 and 383 (Machinery, except electrical, and Machinery, electric). Productivity growth rates are lowest in industry 323 (Leather products) and industry 361 (Pottery, china, earthenware). Most Ricardian models suggest using labor productivity to measure comparative advantage. This is fortunate in that manufacturing output and employment data are among the most avail- able metrics for the broad grouping of countries under study. Labor is typically the only factor of production in Ricardian models, so a natural extension might be total factor productivity that also allows for capital accumulation as well. Unfortunately, capital data at the country-industry level for this sample is too sparse to be of bene…t in a panel study. An earlier version of this paper presents results using output to measure industrial specialization. 4 Export Volumes Columns 6-8 of Tables S1a and S1b describe the WTF data. These descriptives focus on the intensive-margin estimations that require exporter-importer-industry observations maintain the minimum threshold of trade volume. Columns 6 and 7 provide comparable statistics about the mean export levels and growth rates for included routes. Germany and Japan have the highest average volumes, and Nicaragua and El Salvador have the lowest average volumes. Export growth rates are strongest in Nicaragua, Congo, and Costa Rica, and they are lowest in Guatemala and Zimbabwe. From an industry perspective, trade volumes have the highest average values in industries 382-384 (Machinery and Transportation equipment), and the lowest average volumes are observed in industry 361 (Pottery, china, earthenware). Industry 383 (Machinery, electric) has the highest growth rate, while industries 353 (Petroleum re…neries) and 371 (Iron and steel) have the lowest. Column 8 of Table S1a documents the share of total WTF exports for countries that are included in this sample. The main reasons why exports are not included are lack of corresponding UNIDO labor productivity estimates or that the exports are going to the United States. This sample accounts for 78% of exports to destinations other than United States from these countries (about 63% if exports to the United States are included in the denominator). Column 8 of Table S1b provides comparable data for industries. The sample accounts for 69% of exports in these industries to destinations other than United States. This share is lower than 78% due to the inclusion of exporters not captured in Table S1a. Much of the decline on the industry side comes through limited representation of major petroleum producers. Price de‡ators are not available for this sample (exports or labor productivity data). To the extent that exporter-industry-year de‡ators are comprised of exporter-year, industry-year, and exporter-industry components, the …xed e¤ects and …rst di¤erencing strategy will control for them automatically. Residual exporter-industry-year trends could bias OLS estimations. The IV estimations will overcome any such OLS biases due to de‡ators. As a …nal note, a break exists in WTF data collection procedures at 1984. This break does not have a signi…cant impact on ISIC3-level export volumes, and the results are robust to dropping the initial period. 5 US Historical Settlement Patterns s ancestry population. Columns 9 and 10 of Table S1a provide the largest cities for each country’ Due to their large overall size and high immigration shares, cities like New York, Los Angeles, and Miami appear frequently. Nonetheless, there is substantial heterogeneity. East and West Coast cities are more likely to link respectively to European and Asian ancestries, for example, while southern parts of the United States link more to Latin American ancestries. The migration s presence in Minnesota). from Nordic countries to the mid-west is evident (e.g., Finland’ Robustness checks use the spatial distributions present in the 1990 Census of Populations. These patterns are constructed from the 5% sample in a manner that parallels the 1980 distri- bution. US Patenting Data The addresses listed on inventor records are used to group patents to the cities identi…ed in the 1980 census. This procedure uses city lists collected from the O¢ ce of Social and Economic Data Analysis at the University of Missouri, with a matching rate of 99%. Manual recoding further ensures that all patents with more than 100 citations and all city names with more than 100 patents are identi…ed. Some smaller metropolitan areas identi…ed in the patent data are excluded since they do not link to places identi…ed in the 1980 census. Only patents with all inventors living in the United States at the time of their patent application are included, and multiple inventors are discounted so that each patent receives the same weight when measuring inventor populations. Second, the USPTO issues patents by technology categories rather than by industries. The work of Johnson (1999), Silverman (1999), and Kerr (2008) develops concordances that link the USPTO classes to ISIC3 industries in which new inventions are manufactured or used. The main estimations focus on industry-of-use, a¤ording a composite view of the technological opportunity developed for an industry. Studies of advanced economies …nd accounting for these inter-industry R&D ‡ows important (e.g., Scherer 1984, Keller 2002b). Estimations using the alternative categorization of technologies to industry of manufacturer are also presented below. Cohen (2011) discusses the larger literature on industry-level mappings and evidence regarding patents, R&D, and productivity. With respect to ethnic name matching, Kerr (2007) documents speci…c algorithms, lists frequent ethnic names, and provides extensive descriptive statistics. This paper also discusses 6 quality assurance exercises performed. For example, the ethnic-name database can be applied to foreign patents registered with the USPTO. The ethnic-name database assigns ethnicities to 98% of foreign records. Moreover, estimated inventor compositions are quite reasonable— for example, approximately 90% of inventors …ling from Chinese countries and regions are classi…ed as ethnically Chinese. k The developed approach, in combination with the interactions embedded in model Mit out- lined in the main text, has the subtler advantage of allowing the construction of instruments for many more countries and their ancestry groups than what one can directly identify through inventor names, either due to lack of names for some countries (e.g., Ethiopia, Oman) or due to extensive name overlap among countries within an ethnic group (e.g., most of Latin Amer- ica sharing Hispanic names). The instrument also has only non-zero predicted changes for country-industries due to its summation over cities and the connection to Anglo-Saxon technol- ogy development in cities. The Costinot et al. (2012) model dictates the inclusion of importer-industry-year …xed e¤ects k jt in the …rst-di¤erenced speci…cation for structural reasons such as country preferences. Two other rationales exist for having at least industry-year …xed e¤ects due to the data development depicted. First, the US technology frontier is taken as exogenous in the IV model, and the identi…cation of the parameter should thus be independent of the pace of US technology expansion in di¤erent industries. A second methodological rationale stems from the US patent process. US patent grants have increased dramatically since the early 1980s. While several factors lie behind this increase, it is clear that USPTO grant rates grew faster than the underlying growth of US scienti…c personnel and innovation can explain. Moreover, di¤erences in grant rates exist across industries. The …xed e¤ects account for these secular changes in the underlying patenting productivity.1 IV Extended OLS Empirical Results Table S3 considers several extensions of Table 1 to characterize heterogeneity in the sample. Column 2 interacts the regressor with the GDP/capita level of the exporter, broken down into quintiles. Interestingly, the OLS link between base productivity and exports is mainly coming o¤ of the lower-income countries, suggestive of higher trade due to varieties among developed economies. This may be due to the greater potential for within-country productivity dispersion 1 For example, Griliches (1990), Kortum and Lerner (2000), Kim and Marschke (2004), Hall (2005), Branstetter and Ogura (2005), Ja¤e and Lerner (2005), and Lemley and Sampat (2008). 7 in the sample of developing and emerging economies, their more rapid productivity development over the 1980-1999 period, and similar factors. It is similar to the conclusion of Fieler (2011) that trade among advanced economies links to product di¤erentiation and variety (low ), while trade among emerging economies links more closely to fundamental productivity levels (higher ).2 By contrast, Column 3 …nds very little di¤erence across countries of di¤erent sizes. Columns 4 and 5 also show very little connection of export growth to geographic distances, excepting the fact that the growth in exports is not simply happening to bordering countries. These extensions suggest that spatial distance is a second-order factor in shaping where export growth occurs following technology expansion. In contrast to the Ricardian framework, Heckscher-Ohlin-Vanek (HOV) models describe trade as resulting from factor di¤erences across countries (e.g., labor, capital, natural resources). In the above model, technology is the only channel promoting export growth due to identical factor endowments and no intertemporal factor accumulation. During the period studied, some countries experienced signi…cant growth in their skilled labor forces and physical capital stocks, as well as their technology sets, and the former could lead to growth in manufacturing exports due to the Rybczynski e¤ect. Capital accumulation is particularly noted in rapid advances made by several East Asian economies (e.g., Young 1992, 1995; Ventura 1997). The inclusion of exporter-importer-year …xed e¤ects suggests that a Rybczynski e¤ect for the manufacturing sector as a whole is not responsible for the observed trade patterns. Columns 6-8 provide additional evidence that the observed role for technology within manufacturing is not due to specialized factor accumulations.3 The intuition behind the proposed test is straightforward. Under the Rybczynski e¤ect, the s specialization towards manufac- accumulation of skilled workers in country i shifts country i’ turing industries that employ skilled labor more intensively than other factors. By grouping s impor- manufacturing industries by their skilled-labor intensities, tests examine if technology’ tance is preserved after time trends are removed for these industry groups within each country. These time trends are included in addition to the …xed e¤ects listed at the bottom of the table. To illustrate, the computer and pharmaceutical industries are both highly skill intensive. A general Rybczynski e¤ect due to skilled worker accumulation in China would favor specializa- 2 These interactions are an empirical extension that are beyond the closed-form model depicted above. Fieler (2011) provides a theoretical foundation for this work. 3 See Heckscher (1919), Ohlin (1933), and Vanek (1968). Dornbusch et al. (1980) provide a classic HOV model, while Schott (2003) and Romalis (2004) o¤er powerful extensions and empirical tests. Tre‡ er (1994, 1996), Harrigan (1997b), Davis and Weinstein (2001), Chor (2010), Morrow (2010), and Burstein and Vogel (2012) jointly explore technology and factor di¤erences as determinants of trade. 8 s role is tion and export growth in these industries equally. Additional con…dence for technology’ s exports grow faster in the skill-intensive industry that receives the strongest warranted if China’ technology transfer from the United States relative to its peer industries. To implement this matching exercise, industries are grouped into quintiles based upon their s capital-labor factor intensities in the United States. Three intensities are studied— the industry’ s mean wage rate, and the share of non-production workers in the industry’ ratio, the industry’ s labor force. Table S1b documents for each industry the quintile groupings assigned. Textiles and apparel consistently rank in the lowest quintile in all three classi…cations schemes, while chemicals and industrial machinery consistently fall into top quintiles. Some di¤erences do exist though. The correlations among quintile groupings are 0.77 for capital-labor and mean wage, 0.60 for mean wage and non-production share, and 0.44 for capital-labor and non-production share. The role for technology holds up well in all three variants. These …ndings suggest an omitted factor accumulation is not confounding the identi…ed role for technology. The NBER working paper for this article further discusses tests and limitations at this interface (e.g., Leontief 1953, Leamer 1980, Davis and Weinstein 2001, Romalis 2004, Morrow 2010). V Extended IV Empirical Results Extensive Margin Results The reported estimations in the main text focus on the intensive margin of export growth. Table S4 reports OLS and reduced-form results related to the extensive margin of trade com- mencement. These estimations follow Tables 1 and 3 in their approach, with linear probability models considering an indicator variable for exports above US $100,000 on a bilateral route as the outcome variable. The OLS and reduced-form results do not display signi…cant elasticities. Computational limits prevent the full estimation of IV elasticities, but the weak and insigni…cant reduced-form elasticities are su¢ cient to conclude that the IV results would also be weak. On the whole, this study concludes that the exporting growth due to enhanced labor productivity and technology transfer comes through export growth on existing routes rather than through entry into new bilateral routes. Baldwin and Harrigan (2011) identify for the United States some weaknesses of the compar- ative advantage model for accounting for export zeros. The test in this paper links exporting in a speci…c industry with technology for that industry. This approach di¤ers from examinations of the extensive margins of trade that count the number of independent varieties exported (e.g., 9 Feenstra 1994, Feenstra and Rose 2000, Hummels and Klenow 2005). Extended IV Results: Basic Robustness Table S6a provides robustness checks in the same format as Table 2, and Table S6b provides reduced-form relationships. The instrumented elasticity is again very stable to the exclusion of high-pro…le countries, and sample composition is further considered below. The weighted elasticity is very stable in Column 4 to excluding industry 383 (Machinery, electrical), while the unweighted elasticity strengthens. This choice of reporting is again due to electrical machinery having the largest impact on the results, with, for example, very similar elasticities to Column 1 being observed if excluding industry 382 (Machinery, except electrical). Results are also again very similar with a winsorized sample. The instrumented elasticities are reasonably stable to variations on industry dimensions and samples in Columns 6-8 in terms of economic magnitudes, but the standard errors on the unweighted sample in Panel B become too large for de…nitive conclusions. The last column again shows the results are robust to exporter-level clustering. The F statistics with this approach are 8.9 and 11.9 for Panels A and B, respectively. Employing bilateral-route clustering or bootstrapped standard errors substantially increases the precision of the results, with the standard errors declining in size by 60-70%. The F statistics also increase substantially. Bilateral-route clustering is more comparable to prior estimates in the trade literature, but the exporter-industry clustering is a conservative approach common in applied econometrics. The primary IV estimations build a …ve-year lag structure into when the technology growth occurs in the United States to when the productivity growth happens abroad. Thus, growth in technology ‡ows over the 1975-1979!1980-1984 period towards an exporter-industry are used to predict labor productivity growth and export growth during the 1980-1984!1985-1989 period. This …ve-year lag matches discussions of rates of di¤erential technology ‡ows across countries. Table S5 compares this lag structure with contemporaneous technology ‡ows in estimations similar to the …rst-stage and reduced-form speci…cations. The lagged estimator is stronger than the contemporaneous estimator when both are modelled independently or when modelled jointly. This pattern provides comfort in the estimation design and the proposed causal direction of the results. Unreported estimations model additional …xed e¤ects for distance-industry-year, where dis- tance is an indicator variable for being more than the median distance from the United States. 10 This augmented speci…cation controls even more tightly for geographical distance as a determi- nant of technology di¤usion, …nding continued and strong evidence that di¤erential technology transfer from the United States matters. Similar results are also found when including an equiv- alent set of …xed e¤ects that partition on GDP/capita of countries. Additional estimations also test including the three Rybczynski e¤ect controls discussed ear- lier. In the unweighted estimations, the inclusion of these controls does not materially in‡uence the instrumented elasticities. All three elasticities are in the range of 1.5-2.1 and are statistically signi…cant at a 5% level or higher. In the weighted estimations, these controls have a larger im- pact. For the non-production share control, the weighted elasticity is 2.5 (1.7), while estimations with the other two controls have invalid …rst stages. Thus, the conclusions regarding heteroge- neous technology transfer, productivity growth, and exports for the weighted estimations need to be cautious to acknowledge that these e¤ects are not well-distinguished from a generalized Rybczynski e¤ect operating inside of the manufacturing sector itself. In unweighted estimations, these e¤ects are better distinguished. Extended IV Results: Instrument Design Table S7a reports additional robustness checks related to the instrument design, with Table S7b again providing reduced-form results. Columns 2-4 consider variations in the city-industry k technology trend term in Mit outlined in the main text, while Columns 5-9 consider variations in the cross-sectional distributions used to weight US cities for countries. Column 2 begins by showing similar results when using the total technology trends that en- k compass all ethnic groups in model Mit , rather than just the Anglo-Saxon ethnic trend. Column 3 goes the opposite direction by emphasizing even further isolated Anglo-Saxon-led develop- ments. A possible concern for achieving identi…cation with the Anglo-Saxon ethnic technology trend is that it could be very tightly integrated to migrant contributions and their inputs into the innovation process, ranging from the generation of breakthrough ideas to serving as research assistants. While it is impossible to fully disentangle all of these features, patent citations pro- vide useful traction for considering knowledge collaboration and information exchange among peers. Column 3 reduces the emphasis on Anglo-Saxon-invented patents when the patents cite a substantial number of recent patents with non-Anglo-Saxon inventors in the United States; in contrast, it emphasizes even more the Anglo-Saxon-invented patents where the citations are mostly going to recent work by other Anglo-Saxon ethnic inventors.4 These estimations deliver 4 More speci…cally, the procedure …rst isolates the “backwards” citations that each patent makes to patents 11 modestly larger values that are statistically signi…cant. While the larger standard errors are evidence of non-trivial empirical strain from “doubling down” on the Anglo-Saxon exclusion idea, this robustness check is quite valuable for asserting that the measured e¤ects do not dis- play …rst-order evidence of reverse causality, recognizing the limits of empirical tests for ruling out these alternatives. More deviation, however, is observed in Column 4 when using the technology-to-industry concordances that emphasize where manufacturing occurs, rather than where technologies are used. The main estimations focus on industry-of-use, a¤ording a composite view of the techno- logical opportunity developed for an industry. Keller (2002b) reports inter-industry R&D ‡ows aid productivity growth signi…cantly within OECD countries, equal to half or more of the own- industry development. Estimations with manufacturing industries support the using-industry speci…cations in a weighted format, albeit with larger standard errors. Unweighted estimates have a zero-valued …rst stage that prevents further analysis. This di¤erence emphasizes the importance of technology adoption behind the labor productivity results. Columns 5-9 test variations on the construction of M %i;c;1980 . Columns 5 and 6 show that very similar results are obtained when using …rst- or later-generation migrants for de…ning the spatial patterns of migrants. This comparability is not surprising given the persistence of ethnic enclaves in the United States and their attraction of new immigrants from the home country. This stability suggests that the results are not being in‡uenced by endogenous migrant decisions about which cities will show particular strength in patenting growth for certain industries. Columns s educated workers from a country’ 7 and 8 show similar results when using bachelor’ s ancestry s degrees, with the weighted estimations somewhat favoring the versus those without bachelor’ s educated workers. Overall, these variations suggest a strong stability distribution of bachelor’ s construction. Finally, Column 9 shows the results strengthen slightly to this part of the IV’ when using the 1990 city distribution of ancestry stocks for each country. The ancestry variable in the 1990 census is comparable to that in the 1980 census, albeit with slightly less detail. While it is better to focus on the 1980 distributions given that they are fully pre-determined for the sample period, this stability to using 1990 distributions is quite important given the substantial …led over the prior three years. This …xed window uses application years for patents and their citations. A three- year window focuses on recent knowledge development; a …xed window is also important for having a uniform approach over the sample period, as the ethnicity of inventors can only be determined on patents granted after 1975 when the inventor records become digitized. From this citation set, the procedure next estimates the US- based inventor share of cited work that is not of Anglo-Saxon ethnicity. This is a share-based measure given that patents cite multiple prior patents and most patents have multi-inventor teams. The additional weighting applied to patents is one minus this share. This creates e¤ectively a double weight that multiplies the share of the patent’ s own inventor US-based team that is of Anglo-Saxon origin with the added weight based upon the patent’s citations. 12 spatial shifts in economic activity in the United States towards the south and west during this period. Extended IV Results: Sample Composition Tables S8a and S8b consider additional variations on sample composition beyond those reported in Tables S6a and S6b. Columns 2-4 consider the exclusion of countries with limited data points, with Column 3 of Table S1a listing observation counts by country. These tests are a …rst-order check on data quality by considering settings with richer inputs that have been collected over a longer period. Implicitly, the sample focus across Columns 2-4 shifts increasingly towards advanced economies as the requirement moves from 100 data points to 2000 data points. The sample does however retain some important emerging economies like China and India. The coe¢ cients decline somewhat across these columns, but remain quite comparable and precisely measured despite excluding up to 30% of the base sample. Column 5 takes a di¤erent route and includes only long-term OECD countries. This increases the sample size from Column 4, but excludes all emerging economies. The weighted results show a smaller coe¢ cient that is no longer precisely measured, while the unweighted results in Panel B are stronger. Column 6 demonstrates robustness to dropping countries with a limited migrant connection to the United States, speci…cally those nations with fewer than 100,000 people in the United States in 1980 re- porting ancestral connections to the country. These various approaches highlight that the central conclusions of the IV estimations are robust to many settings and that the emerging economies in large part help to sharpen the precision of the estimates. The comparative advantage of this paper remains estimations that include 88 exporters and many destinations, but the results are robust to considering subsets that emphasize data quality features. Columns 7-10 analyze further the concern that productivity advances abroad might result in s specialized the endogenous migration of talent to the United States that is skilled in the country’ work and the simultaneous heightened exports of the country itself. These extra tests in Tables S8a and S8b take the aggressive approach of removing for each nation the top one or …ve cases that would generate the most concern to see if the results still hold. (Of course, the worry with these types of robustness checks is that they exclude industries within each country that are also prime candidates for the technology transfer account outlined in this paper.) In Columns 7-8, industries are excluded on the basis of being the initial top patenting industry abroad for a country; in Columns 9-10, industries are dropped if they are those that show the highest labor 13 productivity for a country adjusted for country and industry …xed e¤ects.5 The results are overall robust to these tests, with the one failure happening in the …rst stage of the weighted analysis that excludes …ve top patenting industries for each country. These exclusions tend to result in slightly higher estimates, but the most important conclusion is that the results are broadly robust to these exclusions of special cases. Extended IV Results: Identi…cation This section concludes with an extended discussion of identi…cation, consolidating some of the earlier results in the paper and presenting one new one. It is very important that the instrument not re‡ect reverse causality or omitted factors— that is, that technology transfer from the United States facilitated by past migration be promoting foreign productivity increases abroad and the concomitant exports, rather than foreign productivity advances leading directly to both exports and heightened migration to the United States that results in greater patenting. This concern …nds its greatest expression when contemplating a foreign multinational like Sony or Mercedes- Benz exporting new products and sending their inventors to the United States to engage in R&D and patenting. An empirical study can never rule out every possible aspect of these inter-relationships, but k the progress here is substantial. The instrument design Mit outlined in the main text holds …xed the 1980 ancestry spatial distribution in the United States for countries and interacts this distribution with Anglo-Saxon-led technology trends by city-industry. Identi…cation with this IV design and the modelled …xed e¤ects in the …rst-di¤erenced speci…cation (3) of the main text thus limits reverse causality to being present in the 1980 distribution, in a way that anticipates s booming industries by US city, or that reverse causality future technology growth for a country’ come through the technology trend itself. For the …rst part about endogenous US city selection being present by 1980, it is important to recall that the city distribution does not itself provide identi…cation given the …rst di¤er- encing and …xed e¤ects included in the …rst-di¤erenced speci…cation. What matters instead s US spatial foot- is the di¤erential exposure across industries and technologies that a country’ s spatial placements by print generates. Thus, for a bias to exist, it must be that a country’ 1980 sought out speci…c city-industry combinations in the United States that would become 5 The top industries for each country in Columns 7-8 are determined through USPTO patent …lings made by inventor teams working in each country and industry during 1975-1984. For Columns 9-10, the procedure regresses country-industry labor productivity on a set of country, industry, and year …xed e¤ects. For each country, the one or …ve industries with the highest average residual labor productivity are excluded. 14 s exports over the ensuing two decades. Given the disproportionately relevant to the country’ importer-industry-year …xed e¤ects, this selection must also be disproportionate for the coun- try compared to other nations. Important tests against this concern include the exclusions of candidate industries in Table S8a and the robustness checks on spatial distributions in Table S7a. With respect to the technology trends of US cities being endogenous, the robustness checks consider a variety of techniques, including the stringent restrictions around technology trends built upon Anglo-Saxon inventors and the work they cite. Another test utilizes the ethnic patenting growth in the United States to formulate an additional control against reverse causality. This control is calculated as the patenting growth within each industry in the United States by members of the focal ethnic community. Thus, estimations consider the technology transfer k instrument Mit outlined in the main text and its impact for productivity growth and trade after controlling for the direct growth of patenting by industry of ethnic communities. The control is calculated across the ethnic groups discussed for ethnic patenting data, with the same control applied to countries within an ethnicity (e.g., the growth in US patenting by ethnic s and Chile’ Hispanic inventors for computers is used as a control with both Mexico’ s computer productivity and exports). As some countries do not map to an identi…able ethnic group with the name matching approach, the sample is reduced to 73,545 observations. For this sample, the weighted and unweighted instrumented elasticities before the ethnic patenting control is introduced are 1.32 (0.54) and 2.21 (0.92), respectively. With the control, these instrumented elasticities are very similar at 1.40 (0.64) and 2.32 (1.37), respectively. These too suggest the technology transfer is happening over and above direct immigrant contributions. Earlier versions of this paper consider direct ethnic patenting and exports more extensively. The early work also evaluates the role of sector reallocation from agriculture, the empirical contrast of the technology states in the exporter and the importer, and the potential for vertical integration through parts trade (e.g., Ng and Yeats 2001, Schott 2004). These results are available upon request. The timing of e¤ects is important as shown in Table S5, where the …ve-year lagged instru- ments outperform contemporaneous variables. There is not a simple story that identi…es why a productivity boom in an Indian industry, for example, would become …rst visible in spillover patenting to US-based Anglo-Saxon inventors versus …rst increasing directly Indian exports in the industry. This timing could be argued— if, for example, local institutions or complementary factors in India are not developed su¢ ciently to realize exports without a signi…cant delay— but the inverted timing structure is far from obvious and would need to be systematic across coun- 15 tries to be picked up in the estimates. On the other hand, the technology transfer story would suggest Anglo-Saxon patenting would precede export growth abroad by India. These types of timing arguments are not perfect, which is why they are relegated to an appendix, but are useful for thinking through the technology trend argument. In the end, the paper is able to make substantial progress towards causal identi…cation in a Ricardian model, beginning with the panel estimation approach and extending through the many IV approaches. This e¤ort remains incomplete, however, and it is hoped that future work both identi…es natural experiment settings to test these arguments and also identi…es other forms/impetuses for heterogeneous technology transfer that can provide identi…cation in a setting that focuses on large country-industry samples like this one. Additional Supplemental References Baldwin, Richard, and James Harrigan. 2011. “Zeros, Quality, and Space: Trade Theory and Trade Evidence.” American Economic Journal: Microeconomics 3 (1): 60-88. Branstetter, Lee, and Yoshiaki Ogura. 2005. “Is Academic Science Driving a Surge in Industrial Innovation? Evidence from Patent Citations.” NBER Working Paper 11561. Cambridge, Mass.: National Bureau of Economic Research. Burstein, Ariel, and Jonathan Vogel. 2012. “Factor Prices and International Trade: A Unifying Perspective.” Working Paper. Coe, David, and Elhanan Helpman. 1995. “International R&D Spillovers.” European Economic Review 39 (5): 859-887. Coe, David, Elhanan Helpman, and Alexander Ho¤maister. 1997. “North-South R&D Spillovers.” Economic Journal 107 (440): 134-149. Cohen, Wesley. 2011. “Fifty Years of Empirical Studies of Innovative Activity and Performance.” In B. Hall and N. Rosenberg, eds., Handbook of the Economics of Innovation. Amsterdam: Elsevier. Dornbusch, Rudiger, Stanley Fischer, and Paul Samuelson. 1980. “Heckscher-Ohlin Trade Theory with a Continuum of Goods.” Quarterly Journal of Economics 95 (2): 203-224. Feenstra, Robert. 1994. “New Product Varieties and the Measurement of International Prices.” American Economic Review 84 (1): 157-177. Frankel, Je¤rey. 1997. Regional Trading Blocs in the World Economic System. Washington, DC: Institute for International Economics. Griliches, Zvi. “Patent Statistics as Economic Indicators: A Survey.” Journal of Economic Literature 28 (4): 1661-1707. Griliches, Zvi. “Productivity, R&D, and the Data Constraint.” American Economic Review 84 (1): 1-23. Grossman, Gene, and Elhanan Helpman. 1991. Innovation and Growth in the Global Economy. Cambridge, Mass.: MIT Press. Hall, Bronwyn. 2005. “Exploring the Patent Explosion.” Journal of Technology Transfer 30 (2.2): 35-48. 16 Harrigan, James. 1997. “Cross-Country Comparisons of Industry Total Factor Productivity: Theory and Evidence.” Federal Reserve Bank of New York Research Paper 9734. Heckscher, Eli. 1919. “The E¤ect of Foreign Trade on the Distribution of Income.” Ekonomisk Tidskrift 21 (2): 1-32. Helpman, Elhanan. 1999. “R&D and Productivity: The International Connection.” In A. Razin and E. Sadka, eds., The Economics of Globalization. Cambridge: Cambridge University Press. Ja¤e, Adam, and Joshua Lerner. 2005. Innovation and Its Discontents. Boston, Mass.: Harvard Business School Press. Ja¤e, Adam, and Manuel Trajtenberg. 1999. “International Knowledge Flows: Evidence from Patent Citations.” Economics of Innovation and New Technology 8: 105-136. Johnson, Daniel. 1999. “150 Years of American Invention: Methodology and a First Geographic Application.” Wellesley College Economics Working Paper 99-01. Kapur, Devesh. 2001. “Diasporas and Technology Transfer.” Journal of Human Development 2 (2): 265-286. Kapur, Devesh, and John McHale. 2005. “Sojourns and Software: Internationally Mobile Human Capital and High-Tech Industry Development in India, Ireland, and Israel.” In A. Arora and A. Gambardella, eds., From Underdogs to Tigers: The Rise and Growth of the Software Industry in Some Emerging Economies. Oxford: Oxford University Press. Keller, Wolfgang. 2002a. “Geographic Localization of International Technology Di¤usion.” American Economic Review 92 (1): 120-142. Keller, Wolfgang. 2002b. “Trade and the Transmission of Technology.” Journal of Economic Growth 7: 5-24. Keller, Wolfgang. 2004. “International Technology Di¤usion.” Journal of Economic Literature 42 (3): 752-782. Kim, Jinyoung, and Gerald Marschke. 2004. “Accounting for the Recent Surge in U.S. Patent- ing: Changes in R&D Expenditures, Patent Yields, and the High Tech Sector.” Economics of Innovation and New Technologies 13 (6): 543-558. Kortum, Samuel, and Joshua Lerner.2000. “Assessing the Contribution of Venture Capital to Innovation.” RAND Journal of Economics 31 (4): 674-692. Kugler, Maurice, and Hillel Rapoport. 2012. “Migration, FDI and the Margins of Trade.” Unpublished Working Paper. Leamer, Edward. “The Leontief Paradox, Reconsidered.” Journal of Political Economy 88 (3): 495-503. Lemley, Mark, and Bhaven Sampat. 2008. “Is the Patent O¢ ce a Rubber Stamp?” Emory Law Journal 58: 181. Leontief, Wassily. 1953. “Domestic Production and Foreign Trade: The American Capital Position Re-examined.” Proceedings of the American Philosophical Society 97 (4): 332-349. Nanda, Ramana, and Tarun Khanna. 2010. “Diasporas and Domestic Entrepreneurs: Evidence from the Indian Software Industry.” Journal of Economics & Management Strategy 19 (4): 991-1012. Ng, Francis, and Alexander Yeats. 2001. “Production Sharing in East Asia: Who Does What for Whom and Why?” In Global Production and Trade in East Asia. New York: Springer US. Obukhova, Elena. 2008. “Global-Pull, State-Push, or Brain Circulation? Firm-Level Techno- s Semiconductor Design Industry.” Working Paper. logical Upgrading in Shanghai’ Obukhova, Elena. 2009. “Does Brain Circulation Promote International Development? High- Skilled Migration and Organizational Performance.” Working Paper. 17 Ohlin, Bertil. 1933. Interregional and International Trade. Cambridge, Mass.: Harvard Univer- sity Press. Papageorgiou, Chris, and Antonio Spilimbergo. 2008. “Learning Abroad and Technology Adop- tion.” Unpublished Working Paper. Romalis, John. 2004. “Factor Proportions and the Structure of Commodity Trade.” American Economic Review 94 (1): 67-97. Scherer, Frederic. “Using Linked Patent Data and R&D Data to Measure Technology Flows.” In Z. Griliches, ed., R & D, Patents and Productivity. Chicago, IL: University of Chicago Press. Schott, Peter. 2003. “One Size Fits All? Heckscher-Ohlin Specialization in Global Production.” American Economic Review 93 (2): 686-708. Schott, Peter. 2004. “Across-Product versus Within-Product Specialization in International Trade.” Quarterly Journal of Economics 119 (2): 647-678. Silverman, Brian. 1999. “Technological Resources and the Direction of Corporate Diversi…ca- tion: Toward an Integration of the Resource-Based View and Transaction Cost Economics.” Management Science 45 (8): 1109-1124. Tre‡ er, Daniel. 1993. “International Factor Price Di¤erences: Leontief was Right!” Journal of Political Economy 101 (6): 961-987. Tre‡er, Daniel. 1995. “The Case of the Missing Trade and Other Mysteries.” American Eco- nomic Review 85 (5): 1029-1046. Vanek, Jaroslav.1968. “The Factor Proportions Theory: The N-Factor Case.” Kyklos 21 (4): 749-756. Ventura, Jaume. 1997. “Growth and Interdependence.” Quarterly Journal of Economics 112 (1): 57-84. Waugh, Michael. 2010. “International Trade and Income Di¤erences.” American Economic Review 100 (5): 2093-2124. Young, Alwyn. 1992. “A Tale of Two Cities: Factor Accumulation and Technical Change in Hong Kong and Singapore.” National Bureau of Economic Research Macroeconomics Annual 7: 13-63. Young, Alwyn. 1995. “The Tyranny of Numbers: Confronting the Statistical Realities of the East Asian Growth Experience.” Quarterly Journal of Economics 110 (3): 641-680. 18 Table S1a: Descriptive statistics for exporting countries Count of UNIDO productivity WTF manufacturing exports Count of intensive- Mean Mean log 5- Mean Share of periods in margin labor prod. year exports per Mean log 5- total WTF Second largest sample as obs. as over growth for 5-year year exports Largest 1980 city 1980 city for exporter exporter included included period growth rate included in for ancestry ancestry Country after FD after FD industries industries included for routes sample population population (1) (2) (3) (4) (5) (6) (7) (8) (9) (10) Afghanistan 1 15 3.7E+03 0.01 2.2E+08 -0.15 0.79 San Francisco, CA Dallas-Ft. Wth, TX Algeria 3 40 3.3E+04 -0.13 2.6E+08 -0.09 0.01 Tampa-St. Pete, FL Boston, MA Argentina 3 819 9.0E+04 0.28 3.3E+10 0.36 0.64 New York, NY Los Angeles, LA Australia 3 1819 1.4E+05 0.19 6.4E+10 0.42 0.68 Los Angeles, LA New York, NY Austria 3 3702 1.4E+05 0.35 1.6E+11 0.53 0.97 New York, NY Los Angeles, LA Bangladesh 3 194 3.7E+03 0.07 3.6E+09 0.54 0.44 New York, NY Los Angeles, LA Barbados 3 18 5.1E+04 0.05 9.9E+07 0.18 0.33 New York, NY Boston, MA Belgium 3 1445 1.4E+05 0.16 9.4E+10 0.45 0.20 Detroit, MI Green Bay, WI Bolivia 3 102 6.8E+04 -0.05 1.3E+09 0.11 0.69 Los Angeles, LA New York, NY Brazil 2 829 7.8E+04 0.47 7.8E+10 0.25 0.60 New York, NY Los Angeles, LA Bulgaria 3 640 2.8E+04 0.10 5.2E+09 0.43 0.48 Chicago, IL Los Angeles, LA Cameroon 3 102 2.0E+04 -0.06 1.7E+09 -0.12 0.51 Washington, DC None Canada 3 3876 1.9E+05 0.17 1.2E+11 0.45 0.98 Boston, MA Los Angeles, LA Chile 3 723 2.2E+05 0.15 2.8E+10 0.45 0.86 New York, NY Los Angeles, LA China 3 2618 1.3E+04 0.18 4.0E+11 1.25 0.82 San Francisco, CA New York, NY Colombia 3 876 5.5E+04 0.10 6.9E+09 0.52 0.42 New York, NY Miami, FL Congo 1 2 4.3E+04 0.47 1.1E+07 1.56 0.03 Los Angeles, LA None Costa Rica 3 63 4.6E+04 0.11 3.2E+08 1.31 0.12 New York, NY Los Angeles, LA Cote d'Ivoire 1 41 5.2E+04 -0.08 8.8E+09 0.35 0.85 Baltimore, MD New York, NY Cuba 1 47 2.0E+04 -0.04 2.4E+09 -0.36 0.61 Miami, FL New York, NY Cyprus 3 319 3.9E+04 0.15 1.6E+09 0.30 0.61 New York, NY San Francisco, CA Denmark 2 1954 1.2E+05 0.18 9.9E+10 0.49 0.72 Los Angeles, LA Salt Lake City, UT Dom. Republic 1 14 3.8E+04 -0.37 5.1E+08 0.74 0.35 New York, NY Miami, FL Ecuador 3 252 4.6E+04 0.02 2.9E+09 0.60 0.70 New York, NY Los Angeles, LA Egypt 3 396 3.1E+04 0.21 5.0E+09 0.13 0.55 New York, NY Los Angeles, LA El Salvador 2 5 4.3E+04 0.02 1.7E+06 -0.02 0.00 Los Angeles, LA San Francisco, CA Ethiopia 1 22 1.0E+04 -0.10 2.6E+08 -0.19 0.18 Chicago, IL Atlanta, GA Fiji 2 31 4.3E+04 -0.01 9.3E+08 0.07 0.60 San Francisco, CA Honolulu, HI Finland 3 2319 1.6E+05 0.39 9.7E+10 0.74 0.79 Duluth-Super., MN Minn.-St. Paul, MN France 3 6100 1.6E+05 0.36 7.3E+11 0.48 0.83 Los Angeles, LA Boston, MA Germany 1 1379 2.1E+05 0.28 1.1E+12 0.33 0.68 Chicago, IL New York, NY Ghana 3 51 1.8E+04 0.02 1.4E+09 0.63 0.36 New York, NY Los Angeles, LA Greece 3 1860 9.9E+04 0.14 3.0E+10 0.36 0.91 New York, NY Chicago, IL Guatemala 3 32 2.1E+04 -0.18 1.1E+08 -0.55 0.04 Los Angeles, LA Chicago, IL Honduras 3 51 2.9E+04 -0.11 2.8E+08 0.75 0.16 New York, NY Los Angeles, LA Hong Kong 3 2453 1.0E+05 0.53 1.4E+11 0.56 0.96 New York, NY Sacramento, CA Hungary 3 1446 2.7E+04 0.17 3.8E+10 0.79 0.85 New York, NY Cleveland, OH Iceland 3 177 1.2E+05 0.23 6.2E+09 0.43 0.92 Los Angeles, LA Seattle, WA India 3 2544 1.8E+04 0.14 6.0E+10 0.72 0.81 New York, NY Chicago, IL Indonesia 3 735 2.0E+04 0.12 5.4E+10 1.10 0.45 New York, NY Los Angeles, LA Source: Author's analysis based on data described in text. Notes: Table provides descriptive statistics on intensive-margin sample. An included observation is the first-difference of values for an exporter-importer-industry-period from the prior period. For the differencing, the exporter's labor productivity at the industry level must be reported in the UNIDO database in both periods and the export volumes in the World Trade Flows database must exceed $100k for both periods. All trade with the United States is excluded. Column 2 reports the number of periods after first differencing that a country is included in, and the third column provides route-level observation counts after first differencing. Columns 4 and 5 provide labor productivity estimates from the UNIDO database. Columns 6-8 report statistics from the World Trade Flows database, with volumes expressed in US dollars. The last two columns document the two largest ethnic heritage cities used in the instrument design. Cities are defined at the consolidated metropolitan area level, with abbreviated names provided. Table S1a, continued Count of UNIDO productivity WTF manufacturing exports Count of intensive- Mean Mean log 5- Mean Share of periods in margin labor prod. year exports per Mean log 5- total WTF Second largest sample as obs. as over growth for 5-year year exports Largest 1980 city 1980 city for exporter exporter included included period growth rate included in for ancestry ancestry Country after FD after FD industries industries included for routes sample population population (1) (2) (3) (4) (5) (6) (7) (8) (9) (10) Iran 2 157 2.1E+04 -0.09 4.4E+09 0.43 0.55 Los Angeles, LA San Francisco, CA Ireland 3 2061 2.9E+05 0.41 1.2E+11 0.79 0.96 New York, NY Philadelphia, PA Israel 3 1123 1.0E+05 0.22 1.3E+10 0.69 0.50 New York, NY Los Angeles, LA Italy 3 6102 1.5E+05 0.35 6.2E+11 0.50 0.89 New York, NY Philadelphia, PA Jamaica 2 3 1.1E+05 0.07 7.6E+06 0.92 0.01 New York, NY Miami, FL Japan 3 5799 2.7E+05 0.37 1.2E+12 0.43 0.99 Honolulu, HI Los Angeles, LA Jordan 3 113 5.3E+04 0.20 5.5E+08 0.25 0.26 New York, NY Detroit, MI Kenya 3 231 4.8E+04 0.13 1.1E+09 0.32 0.28 Washington, DC San Francisco, CA Kuwait 3 391 7.8E+05 0.12 1.5E+10 -0.35 0.74 Orlando, FL Providence, RI Macau 3 166 2.7E+04 0.33 3.6E+09 0.34 0.72 Los Angeles, LA Seattle, WA Malaysia 3 1317 6.6E+04 0.33 1.1E+11 1.28 0.61 San Francisco, CA Los Angeles, LA Malta 3 223 2.0E+05 0.50 3.5E+09 0.83 0.68 Detroit, MI New York, NY Mexico 3 1474 7.3E+04 0.21 3.8E+10 0.84 0.82 Los Angeles, LA San Antonio, TX Morocco 3 551 4.2E+04 0.00 1.8E+10 0.54 0.86 New York, NY Chicago, IL Myanmar 2 40 2.4E+03 0.64 3.2E+08 0.38 0.24 Washington, DC Los Angeles, LA Nepal 1 2 7.0E+03 0.21 1.7E+06 0.42 0.00 Boston, MA St. Louis, MO Netherlands 3 5429 2.7E+05 0.29 5.1E+11 0.46 0.91 Los Angeles, LA Grand Rapids, MI New Zealand 3 735 1.0E+05 0.31 1.7E+10 0.40 0.49 Washington, DC San Francisco, CA Nicaragua 1 1 4.1E+04 0.40 7.4E+05 1.90 0.00 Los Angeles, LA San Francisco, CA Nigeria 3 28 1.1E+05 -0.06 7.9E+07 0.04 0.03 New York, NY Washington, DC Norway 3 2295 3.4E+05 0.26 8.5E+10 0.36 0.90 Minn.-St. Paul, MN Seattle, WA Oman 1 14 3.1E+05 -0.19 1.4E+09 0.60 0.33 Los Angeles, LA Salinas, CA Pakistan 3 784 2.1E+04 0.19 2.1E+10 0.46 0.87 New York, NY Los Angeles, LA Panama 3 288 3.7E+04 -0.06 3.0E+09 0.07 0.47 New York, NY Los Angeles, LA Peru 3 354 6.6E+04 0.54 3.0E+09 0.26 0.27 New York, NY Los Angeles, LA Philippines 3 930 2.8E+04 0.35 3.5E+10 0.97 0.80 San Francisco, CA Los Angeles, LA Poland 3 1213 2.6E+04 0.25 4.1E+10 0.69 0.71 New York, NY Chicago, IL Portugal 3 2160 6.1E+04 0.39 7.2E+10 0.72 0.96 Providence, RI San Francisco, CA Romania 1 191 6.7E+03 -0.31 8.0E+09 -0.08 0.29 New York, NY Los Angeles, LA Senegal 3 64 3.5E+04 0.10 1.6E+09 0.19 0.81 Los Angeles, LA Washington, DC Singapore 3 2790 2.6E+05 0.51 2.0E+11 0.83 0.94 New York, NY San Francisco, CA South Africa 3 1017 5.1E+04 -0.02 3.1E+10 0.36 0.67 New York, NY Dallas-Ft. Wth, TX South Korea 3 3204 1.2E+05 0.52 2.5E+11 0.95 0.84 Los Angeles, LA New York, NY Spain 3 5130 1.4E+05 0.36 2.6E+11 0.69 0.97 New York, NY Los Angeles, LA Sri Lanka 3 348 8.0E+03 0.20 3.5E+09 0.63 0.59 Los Angeles, LA Boston, MA Sweden 3 4032 1.5E+05 0.30 2.5E+11 0.45 0.98 Minn.-St. Paul, MN Chicago, IL Switzerland 2 908 1.9E+05 0.29 1.1E+11 0.35 0.42 Los Angeles, LA New York, NY Syria 3 87 9.9E+04 0.54 9.1E+08 0.47 0.36 New York, NY Chicago, IL Taiwan 3 2022 7.7E+04 0.46 2.1E+11 0.87 0.74 Los Angeles, LA San Francisco, CA Tanzania 3 54 6.6E+03 -0.35 1.6E+08 0.55 0.09 New York, NY Chicago, IL Thailand 2 499 5.7E+04 0.37 4.2E+10 1.33 0.37 Los Angeles, LA New York, NY Trinidad-Tobago 3 41 1.7E+05 0.11 4.9E+08 -0.36 0.22 New York, NY Washington, DC Tunisia 1 46 8.8E+04 0.25 2.6E+09 -0.39 0.15 New York, NY Los Angeles, LA Turkey 3 1233 6.8E+04 0.25 4.8E+10 0.81 0.80 New York, NY Los Angeles, LA United Kingdom 3 7143 1.6E+05 0.37 6.1E+11 0.48 0.90 Los Angeles, LA New York, NY Uruguay 3 431 4.4E+04 0.20 6.5E+09 0.21 0.88 New York, NY Washington, DC Venezuela 3 490 3.1E+05 -0.03 8.4E+09 -0.01 0.51 New York, NY Miami, FL Zimbabwe 3 14 3.7E+04 0.12 5.7E+07 -0.58 0.01 Los Angeles, LA None Table S1b: Descriptive statistics for ISIC Revision 2 industries UNIDO productivity WTF manufacturing exports US patenting US quintiles (5 = Highest) Count of Mean log Mean log Share of intensive- 5-year Mean 5-year total WTF Share of Mean log Non- margin Mean level growth exports per growth exports patenting 5-year Capital/ prod. obs. after per 5-year rate for 5-year rate for included in included in growth labor Mean worker ISIC Industry title FD period industries period routes sample sample rate ratio wages share (1) (2) (3) (4) (5) (6) (7) (8) (9) (10) (11) (12) (13) 311 Food products 8183 1.5E+05 0.16 5.9E+11 0.33 0.74 0.02 0.15 3 2 3 313 Beverages 2294 2.2E+05 0.31 6.5E+10 0.49 0.75 0.01 0.17 5 3 5 314 Tobacco 1498 3.3E+05 0.22 2.6E+10 0.49 0.56 0.00 0.15 5 5 3 321 Textiles 7067 6.8E+04 0.22 4.9E+11 0.48 0.79 0.02 0.16 2 1 1 322 Wearing apparel, except footwear 3738 4.5E+04 0.18 2.8E+11 0.48 0.83 0.00 0.20 1 1 1 323 Leather products 2605 7.8E+04 0.14 7.8E+10 0.64 0.73 0.00 0.19 1 1 1 324 Footwear, except rubber or plastic 1848 5.8E+04 0.30 4.8E+10 0.42 0.44 0.00 0.19 1 1 1 331 Wood products, except furniture 3217 7.5E+04 0.15 1.1E+11 0.47 0.73 0.01 0.15 2 1 1 332 Furniture, except metal 2551 7.9E+04 0.21 7.9E+10 0.56 0.68 0.01 0.16 1 1 2 341 Paper and products 4382 1.7E+05 0.19 2.0E+11 0.35 0.76 0.03 0.15 4 4 2 342 Printing and publishing 2736 9.6E+04 0.20 4.4E+10 0.43 0.68 0.01 0.16 2 2 5 351 Industrial chemicals 6309 2.5E+05 0.23 5.3E+11 0.46 0.59 0.05 0.10 5 5 4 352 Other chemicals 5379 1.7E+05 0.21 2.8E+11 0.60 0.53 0.08 0.18 4 4 5 353 Petroleum refineries 2257 1.6E+06 0.24 2.2E+11 0.03 0.50 0.01 0.09 5 5 4 354 Misc. petroleum and coal products 722 2.7E+05 0.15 1.8E+10 0.62 0.29 0.01 0.10 4 4 4 355 Rubber products 3819 1.0E+05 0.29 7.8E+10 0.45 0.75 0.01 0.14 3 3 2 361 Pottery, china, earthenware 1648 5.4E+04 0.12 1.4E+10 0.44 0.58 0.00 0.14 1 2 2 362 Glass and products 2750 1.1E+05 0.24 3.7E+10 0.47 0.58 0.01 0.14 4 3 1 369 Other non-metallic mineral products 3157 1.2E+05 0.22 6.6E+10 0.38 0.70 0.01 0.12 4 3 3 371 Iron and steel 4301 2.0E+05 0.29 2.6E+11 0.27 0.59 0.01 0.13 5 5 2 372 Non-ferrous metals 3355 2.1E+05 0.21 1.7E+11 0.35 0.60 0.02 0.16 4 3 3 381 Fabricated metal products 5612 9.0E+04 0.20 2.2E+11 0.42 0.69 0.05 0.15 2 3 3 382 Machinery, except electrical 6989 1.5E+05 0.43 1.1E+12 0.67 0.67 0.25 0.25 3 4 4 383 Machinery, electric 6513 1.5E+05 0.46 1.1E+12 0.74 0.81 0.22 0.26 3 4 5 384 Transport equipment 5714 2.0E+05 0.33 1.0E+12 0.49 0.76 0.09 0.16 2 5 4 385 Professional & scientific equipment 5195 1.1E+05 0.28 2.8E+11 0.45 0.81 0.08 0.24 3 4 5 Source: Author's analysis based on data described in text. Notes: See Table S1a. Column 3 provides route-level observation counts by industry after first differencing. Columns 4 and 5 provide labor productivity estimates from the UNIDO database. Columns 6-8 report statistics from the World Trade Flows database, with volumes expressed in US dollars. Industries 356 and 390 are excluded. Columns 9-10 report the share and growth rate of patenting in United States that is used for the technology transfer measures. The last columns report the quintile to which the industry is assigned for the Rybczynski effect controls of country time trends x industry quintiles. Table S2: Descriptive statistics for inventors residing in United States Ethnicity of inventor Anglo-Saxon Chinese European Hispanic Indian Japanese Korean Russian Vietnam. A. Ethnic Inventor Shares Estimated from US Inventor Records, 1975-2004 (%) 1975-1979 74.8 2.1 15.6 2.7 2.0 0.6 0.3 1.9 0.1 1980-1984 73.4 2.9 15.1 2.7 2.6 0.7 0.4 2.0 0.1 1985-1989 72.2 3.6 14.6 2.9 3.1 0.8 0.5 2.1 0.2 1990-1994 70.0 4.8 14.1 3.2 3.9 0.9 0.6 2.2 0.4 1995-1999 66.4 6.7 13.6 3.5 5.2 0.9 0.7 2.5 0.5 2000-2004 63.1 8.8 13.0 3.8 5.9 1.0 0.9 2.8 0.6 Chemicals 65.8 7.3 14.4 3.2 4.9 0.9 0.7 2.5 0.3 Computers 62.9 8.4 12.6 3.4 7.5 1.0 0.7 2.7 0.7 Pharmaceuticals 64.8 7.2 14.8 3.9 4.6 1.1 0.8 2.6 0.3 Electrical 64.3 8.3 13.3 3.3 5.3 1.0 0.9 2.8 0.7 Mechanical 72.8 3.3 14.2 3.3 2.8 0.7 0.5 2.2 0.2 Miscellaneous 74.1 2.9 13.9 3.6 2.3 0.6 0.5 1.9 0.2 Top Cities as a WS (84) SF (14) MIL (21) MIA (16) SF (8) SD (2) BAL (2) NYC (4) AUS (2) Percentage of SLC (83) LA (8) NOR (19) SA (9) AUS (7) SF (2) LA (1) BOS (4) SF (1) City’s Patents NAS (82) AUS (6) STL (19) WPB (6) PRT (6) LA (2) DC (1) HRT (4) LA (1) B. Immigrant Scientist and Engineer Shares Estimated from 1990 US Census Records (%) Bachelor's Share 87.6 2.7 2.3 2.4 2.3 0.6 0.5 0.4 1.2 Master's Share 78.9 6.7 3.4 2.2 5.4 0.9 0.7 0.8 1 Doctorate Share 71.2 13.2 4 1.7 6.5 0.9 1.5 0.5 0.4 Source: Author's analysis based on data described in text. Notes: Panel A presents descriptive statistics for inventors residing in the US at the time of patent application. Inventor ethnicities are estimated through inventors' names using techniques described in the text. Patents are grouped by application years and major technology fields. Cities, defined through Metropolitan Statistical Areas, include AUS (Austin), BAL (Baltimore), BOS (Boston), DC (Washington), HRT (Hartford), LA (Los Angeles), MIA (Miami), MIL (Milwaukee), NAS (Nashville), NOR (New Orleans), NYC (New York City), PRT (Portland), SA (San Antonio), SD (San Diego), SF (San Francisco), SLC (Salt Lake City), STL (St. Louis), WPB (West Palm Beach), and WS (Winston-Salem). Cities are identified from inventors' city names using city lists collected from the Office of Social and Economic Data Analysis at the University of Missouri, with a matching rate of 99%. Manual recoding further ensures all patents with more than 100 citations and all city names with more than 100 patents are identified. Panel B presents comparable statistics calculated from the 1990 Census using country of birth for scientists and engineers. Anglo-Saxon provides a residual in the Census statistics. Many US inventors with European names are native citizens. Table S3: Robustness checks on OLS specifications in Table 1 Base Including Including Including Including Rybczynski effect controls of estimation Including exporter route border country time trends x industry quintiles (Column 2, GDP/capita populations distance effect Capital/ labor Non-prod. Table 1) interactions interactions interactions interaction ratio Mean wages share (1) (2) (3) (4) (5) (6) (7) (8) The dependent variable is Δ log bilateral exports on the intensive margin by exporter-importer-industry Δ Log country-industry labor 0.573*** 0.938*** 0.578*** 0.269** 0.602*** 0.465*** 0.332** 0.313** productivity (0.185) (0.219) (0.126) (0.125) (0.186) (0.129) (0.135) (0.132) x Second quartile of trait -0.582** 0.272 0.330** indicated in column header (0.283) (0.284) (0.139) x Third quartile of trait -0.972*** -0.368 0.570** indicated in column header (0.271) (0.248) (0.231) x Highest quartile of trait -0.734*** 0.060 0.268 indicated in column header (0.234) (0.269) (0.183) x Bordering economies -0.581*** (0.177) Effect at first quartile 0.938 0.578 0.269 Effect at second quartile 0.356 0.850 0.599 Effect at third quartile -0.034 0.210 0.839 Effect at highest quartile 0.204 0.638 0.536 Value at second quartile start 8,431 8,852,235 2,319 Value at third quartile start 14,765 29,900,000 5,596 Value at highest quartile start 19,024 57,100,000 9,184 Observations 103,839 103,839 103,839 103,839 103,839 103,839 103,839 103,839 Importer-Industry-Yr FE Yes Yes Yes Yes Yes Yes Yes Yes Exporter-Importer-Yr FE Yes Yes Yes Yes Yes Yes Yes Yes Source: Author's analysis based on data described in text. Notes: See Table 1. Interactions in Column 4 use Great Circle distances between capital cities. To give a feel for these demarcations, the distances from Beijing, China, to the capitals of Bangladesh, United Arab Emirates, and Spain are 3029 km., 5967 km., and 9229 km., respectively. Columns 6-8 test for the Rybczynski effect within manufacturing. Industries are grouped into quintiles by their US capital-labor ratios, mean wage rates, and non-production worker wage bill shares. Table S1b lists industry groupings. Linear time trends for each country by industry quintile are included in the estimation. Table S4: Estimations of extensive margin OLS estimation Reduced-form estimation (1) (2) Dependent variable is Δ (0,1) [exports > US$100k] Panel A: Weighting bilateral routes by the interaction of exporter and importer trade in industry Δ Log country-industry labor -0.001 productivity (0.004) Δ Log estimator for technology -0.050 flows from the United States (0.034) Observations 241,792 241,792 Importer-Industry-Yr FE Yes Yes Exporter-Importer-Yr FE Yes Yes Panel B: Excluding sample weights Δ Log country-industry labor -0.002 productivity (0.004) Δ Log estimator for technology -0.023 flows from the United States (0.027) Observations 241,792 241,792 Importer-Industry-Yr FE Yes Yes Exporter-Importer-Yr FE Yes Yes Source: Author's analysis based on data described in text. Notes: See Tables 1 and 3. Estimations test the extensive margin of trade through linear probability models. The dependent variable is a dichotomous indicator variable taking unit value if bilateral exports exceed $100k. The $100k threshold is chosen due to WTF data collection procedures discussed in the text. Table S5: Lag structure of first-stage and reduced-form estimations First-stage estimation Reduced-form estimation (1) (2) (3) (4) (5) (6) Panel A: Weighting bilateral routes by the interaction of exporter and importer trade in industry DV: Δ Log country-industry labor productivity DV: Δ Log bilateral exports Δ Log estimator for technology 0.589** 0.524** 0.938*** 0.886*** flows from the United States, (0.272) (0.241) (0.298) (0.272) lagged five years Δ Log estimator for technology 0.376* 0.275 0.393 0.223 flows from the United States, (0.212) (0.169) (0.253) (0.256) contemporaneous Observations 103,839 103,839 103,839 103,839 103,839 103,839 Importer-Industry-Yr FE Yes Yes Yes Yes Yes Yes Exporter-Importer-Yr FE Yes Yes Yes Yes Yes Yes Panel B: Excluding sample weights DV: Δ Log country-industry labor productivity DV: Δ Log bilateral exports Δ Log estimator for technology 0.267*** 0.265*** 0.648*** 0.636*** flows from the United States, (0.078) (0.078) (0.112) (0.107) lagged five years Δ Log estimator for technology 0.064 0.059 0.473*** 0.459*** flows from the United States, (0.071) (0.069) (0.111) (0.105) contemporaneous Observations 103,839 103,839 103,839 103,839 103,839 103,839 Importer-Industry-Yr FE Yes Yes Yes Yes Yes Yes Exporter-Importer-Yr FE Yes Yes Yes Yes Yes Yes Source: Author's analysis based on data described in text. Notes: See Table 3. The first-stage and reduced-form performance of the lagged technology flows, the preferred instrument, is contrasted with contemporaneous flows. Table S6a: Basic robustness checks on IV specifications Kerr (2008) Kerr (2008) Base Using a Using ISIC sample sample Using estimation Excluding Excluding Excluding 2%/98% 2-digit level using Imp- using Imp- exporter- (Column 3, exports exports electrical winsorized industry ISIC2-Year ISIC3-Year level Table 3) from Brazil from China machinery sample groups fixed effects fixed effects clustering (1) (2) (3) (4) (5) (6) (7) (8) (9) The dependent variable is Δ log bilateral exports on the intensive margin by exporter-importer-industry Panel A: Weighting bilateral routes by the interaction of exporter and importer trade in industry Δ Log country-industry labor 1.592** 1.616** 1.372** 1.616** 1.715** 0.673* 1.109*** 1.794** 1.592*** productivity (0.637) (0.644) (0.649) (0.746) (0.717) (0.393) (0.390) (0.770) (0.335) Observations 103,839 103,010 101,221 97,326 103,839 51,483 23,345 23,345 103,839 Importer-Industry-Yr FE Yes Yes Yes Yes Yes Yes Yes Yes Yes Exporter-Importer-Yr FE Yes Yes Yes Yes Yes Yes Yes Yes Yes Panel B: Excluding sample weights Δ Log country-industry labor 2.429*** 2.485*** 2.331*** 3.415** 2.862*** 1.578 3.448* -0.808 2.429*** productivity (0.791) (0.801) (0.778) (1.735) (0.851) (1.452) (2.059) (2.094) (0.646) Observations 103,839 103,010 101,221 97,326 103,839 51,483 23,345 23,345 103,839 Importer-Industry-Yr FE Yes Yes Yes Yes Yes Yes Yes Yes Yes Exporter-Importer-Yr FE Yes Yes Yes Yes Yes Yes Yes Yes Yes Source: Author's analysis based on data described in text. Notes: See Table 3. Table S6b: Reduced forms for Table S6a Kerr (2008) Kerr (2008) Base Using a Using ISIC sample sample Using estimation Excluding Excluding Excluding 2%/98% 2-digit level using Imp- using Imp- exporter- (Column 2, exports exports electrical winsorized industry ISIC2-Year ISIC3-Year level Table 3) from Brazil from China machinery sample groups fixed effects fixed effects clustering (1) (2) (3) (4) (5) (6) (7) (8) (9) The dependent variable is Δ log bilateral exports on the intensive margin by exporter-importer-industry Panel A: Weighting bilateral routes by the interaction of exporter and importer trade in industry Δ Log estimator for tech 0.938*** 0.956*** 0.698*** 0.748*** 1.044*** 0.597 0.948*** 0.794** 0.938** flows from the United States (0.298) (0.299) (0.220) (0.288) (0.304) (0.387) (0.265) (0.325) (0.434) Observations 103,839 103,010 101,221 97,326 103,839 51,483 23,345 23,345 103,839 Importer-Industry-Yr FE Yes Yes Yes Yes Yes Yes Yes Yes Yes Exporter-Importer-Yr FE Yes Yes Yes Yes Yes Yes Yes Yes Yes Panel B: Excluding sample weights Δ Log estimator for tech 0.648*** 0.664*** 0.602*** 0.545*** 0.764*** 0.259* 0.583*** -0.079 0.648*** flows from the United States (0.112) (0.112) (0.110) (0.110) (0.123) (0.141) (0.180) (0.200) (0.177) Observations 103,839 103,010 101,221 97,326 103,839 51,483 23,345 23,345 103,839 Importer-Industry-Yr FE Yes Yes Yes Yes Yes Yes Yes Yes Yes Exporter-Importer-Yr FE Yes Yes Yes Yes Yes Yes Yes Yes Yes Source: Author's analysis based on data described in text. Notes: See Table 3. Table S7a: Variations in IV design structure Variations in city-industry trend term Variations in city cross-sectional term Using total Weighting Using Using 1990 Base technology patents by industry Using Using non- ancestry estimation trend with backwards groupings Using first- Using later- bachelor's bachelor's distribution (Column 3, all ethnic Anglo- based on generation generation educated educated for city Table 3) groups Saxon cites mfg roles immigrants immigrants workers workers weights (1) (2) (3) (4) (5) (6) (7) (8) (9) The dependent variable is Δ log bilateral exports on the intensive margin by exporter-importer-industry Panel A: Weighting bilateral routes by the interaction of exporter and importer trade in industry Δ Log country-industry labor 1.592** 1.513*** 2.812** 2.634* 1.581*** 1.699 2.233** 1.607** 1.888** productivity (0.637) (0.498) (1.118) (1.583) (0.420) (1.097) (1.052) (0.683) (0.815) Observations 103,839 103,839 103,839 103,839 103,839 103,839 103,839 103,839 103,839 Importer-Industry-Yr FE Yes Yes Yes Yes Yes Yes Yes Yes Yes Exporter-Importer-Yr FE Yes Yes Yes Yes Yes Yes Yes Yes Yes Panel B: Excluding sample weights Δ Log country-industry labor 2.429*** 3.342*** 3.085** Invalid 2.645*** 2.341** 2.727*** 2.812*** 2.650*** productivity (0.791) (1.096) (1.538) first stage (0.733) (1.062) (0.988) (0.962) (0.825) Observations 103,839 103,839 103,839 103,839 103,839 103,839 103,839 103,839 103,839 Importer-Industry-Yr FE Yes Yes Yes Yes Yes Yes Yes Yes Yes Exporter-Importer-Yr FE Yes Yes Yes Yes Yes Yes Yes Yes Yes Source: Author's analysis based on data described in text. Notes: See Table 3. Table S7b: Reduced forms for Table S7a Variations in city-industry trend term Variations in city cross-sectional term Using total Weighting Using Using 1990 Base technology patents by industry Using Using non- ancestry estimation trend with backwards groupings Using first- Using later- bachelor's bachelor's distribution (Column 2, all ethnic Anglo- based on generation generation educated educated for city Table 3) groups Saxon cites mfg roles immigrants immigrants workers workers weights (1) (2) (3) (4) (5) (6) (7) (8) (9) The dependent variable is Δ log bilateral exports on the intensive margin by exporter-importer-industry Panel A: Weighting bilateral routes by the interaction of exporter and importer trade in industry Δ Log estimator for tech 0.938*** 0.999*** 0.795*** 0.248** 1.257*** 0.570* 1.224*** 0.893*** 1.126*** flows from the United States (0.298) (0.311) (0.158) (0.111) (0.387) (0.269) (0.323) (0.300) (0.314) Observations 103,839 103,839 103,839 103,839 103,839 103,839 103,839 103,839 103,839 Importer-Industry-Yr FE Yes Yes Yes Yes Yes Yes Yes Yes Yes Exporter-Importer-Yr FE Yes Yes Yes Yes Yes Yes Yes Yes Yes Panel B: Excluding sample weights Δ Log estimator for tech 0.648*** 0.789*** 0.374*** 0.124*** 0.875*** 0.386*** 0.637*** 0.650*** 0.766*** flows from the United States (0.112) (0.107) (0.063) (0.044) (0.144) (0.102) (0.118) (0.105) (0.120) Observations 103,839 103,839 103,839 103,839 103,839 103,839 103,839 103,839 103,839 Importer-Industry-Yr FE Yes Yes Yes Yes Yes Yes Yes Yes Yes Exporter-Importer-Yr FE Yes Yes Yes Yes Yes Yes Yes Yes Yes Source: Author's analysis based on data described in text. Notes: See Table 3. Table S8a: Variations in IV sample composition Excluding Excluding Excluding Excluding Excluding Excluding 5 Excluding Excluding 5 Base nations with nations with nations with Examining exporters largest largest largest largest estimation fewer than fewer than fewer than long-term with <100k foreign foreign foreign foreign (Column 3, 100 data 1000 data 2000 data OECD ethnic US patenting patenting productivity productivity Table 3) points points points nations only members industry industries outlier outliers (1) (2) (3) (4) (5) (6) (7) (8) (9) (10) The dependent variable is Δ log bilateral exports on the intensive margin by exporter-importer-industry Panel A: Weighting bilateral routes by the interaction of exporter and importer trade in industry Δ Log country-industry labor 1.592** 1.590** 1.470** 1.417** 1.322 1.669** 2.023* Invalid 1.520** 2.563* productivity (0.637) (0.636) (0.568) (0.614) (0.816) (0.797) (1.098) first stage (0.586) (1.554) Observations 103,839 103,013 89,059 71,779 74,087 78,411 97,011 73,518 100,740 82,939 Importer-Industry-Yr FE Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Exporter-Importer-Yr FE Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Panel B: Excluding sample weights Δ Log country-industry labor 2.429*** 2.426*** 2.315*** 1.900*** 2.205** 2.251*** 3.072*** 3.257* 2.853*** 2.923*** productivity (0.791) (0.787) (0.764) (0.671) (0.874) (0.875) (1.179) (1.843) (0.931) (1.081) Observations 103,839 103,013 89,059 71,779 74,087 78,411 97,011 73,647 100,740 82,939 Importer-Industry-Yr FE Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Exporter-Importer-Yr FE Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Source: Author's analysis based on data described in text. Notes: See Table 3. Table S8b: Reduced forms for Table S8a Excluding Excluding Excluding Excluding Excluding Excluding 5 Excluding Excluding 5 Base nations with nations with nations with Examining exporters largest largest largest largest estimation fewer than fewer than fewer than long-term with <100k foreign foreign foreign foreign (Column 2, 100 data 1000 data 2000 data OECD ethnic US patenting patenting productivity productivity Table 3) points points points nations only members industry industries outlier outliers (1) (2) (3) (4) (5) (6) (7) (8) (9) (10) The dependent variable is Δ log bilateral exports on the intensive margin by exporter-importer-industry Panel A: Weighting bilateral routes by the interaction of exporter and importer trade in industry Δ Log estimator for tech 0.938*** 0.937*** 0.936*** 0.904*** 0.622*** 0.975*** 0.898*** 0.716*** 0.948*** 0.747*** flows from the United States (0.298) (0.298) (0.308) (0.341) (0.239) (0.352) (0.175) (0.249) (0.302) (0.274) Observations 103,839 103,013 89,059 71,779 74,087 78,411 97,011 73,518 100,740 82,939 Importer-Industry-Yr FE Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Exporter-Importer-Yr FE Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Panel B: Excluding sample weights Δ Log estimator for tech 0.648*** 0.653*** 0.668*** 0.634*** 0.580*** 0.601*** 0.657*** 0.529*** 0.655*** 0.619*** flows from the United States (0.112) (0.113) (0.120) (0.135) (0.124) (0.124) (0.115) (0.124) (0.116) (0.126) Observations 103,839 103,013 89,059 71,779 74,087 78,411 97,011 73,647 100,740 82,939 Importer-Industry-Yr FE Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Exporter-Importer-Yr FE Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Source: Author's analysis based on data described in text. Notes: See Table 3.