Talent, Geography, and Offshore R&D*

Jingting Fan†

This draft: June, 2020

Abstract

I model and quantify the impact of a new dimension of globalization: offshore R&D. Inthe model, firms work with researchers to develop new product blueprints and then engagein offshore production and exporting. Cross-country differences in the distributions of firmknowhow and worker ability generate a ‘talent-acquisition’ motive for offshore R&D, whilefrictions impeding trade and separation of production from R&D lead to a ‘market-access’ mo-tive. Offshore R&D, production, and trade form a value chain within multinational corpora-tions (MNCs) that facilitates specialization of countries. I discipline the model using empiricalregularities documented from a new firm-level dataset. Counterfactual experiments show thatthe two motives can account for most of the observed offshore R&D. Incorporating offshoreR&D amplifies the gains from globalization by a factor of 1.3 and shapes the gains from tradi-tional forms of global integration, namely trade and offshore production.Keywords: Multinational firms, offshore R&D, global value chain, gains from opennessJEL Classification: F21 F23 F40 O32

*This paper is based on a chapter of my dissertation. I am deeply grateful to Nuno Limao for his encouragementand guidance throughout graduate school and to my committee Sebnem Kalemli-Özcan and John Shea for their in-valuable feedback. I thank Costas Arkolakis, Ariel Burstein, Kamran Bilir, Jonathan Eaton, Gordon Hanson, WolfgangKeller, Eunhee Lee, Wenlan Luo, James Markusen, Luca Opromolla, Natalia Ramondo, Andres Rodriguez-Clare, FelipeSaffie, Lixin Tang, Jim Tybout, Daniel Wilson, Stephen Yeaple, and participants at seminars and conferences for helpfulcomments. All errors are mine.

†Department of Economics, Penn State University. Email address: jxf524@psu.edu

1

1 Introduction

A common narrative of the recent wave of globalization is that reductions in trade barriers andcommunication costs enabled firms from developed countries to offshore production tasks, oftenconsidered of ‘low value-added,’ and to specialize at home in ‘high value-added,’ skill-intensivetasks such management, R&D, and marketing. In line with this view, the literature on multina-tional production and trade has largely focused on the causes and consequences of the allocationof production—or different stages of production—across countries. This narrative, however, cap-tures only partially the complex strategies used by modern firms in organizing activities aroundthe globe. Increasingly, firms engage not only in production but also in R&D in their overseasaffiliates.1 In fact, as Figure 1 shows, after a period of growth, R&D carried out by foreign firmsnow amounts to a substantial share of domestic R&D in many host countries.

The offshore R&D decisions of multinationals corporations (MNCs) could have important ag-gregate implications. By determining the location and efficiency of R&D, it directly affects theincome of countries. Moreover, in a world interconnected through trade and offshore production,by shaping countries’ specialization in innovation or production, it can affect income indirectly.

In this paper I model and quantify the impact of offshore R&D. I address three questions. First,what are the determinants of offshore R&D? Second, how large are the welfare gains of openingup to offshore R&D? Third, how do these gains depend on and interact with the traditional formsof economic integration, namely trade and offshore production?

Answering these questions inevitably require taking a stand on the nature of the connectionbetween R&D and production decisions within an MNC. In particular, assuming that R&D atan affiliate is solely for the purpose of local production can lead to very different answers fromassuming all R&D is for production at the headquarters. Yet the purpose of R&D is rarely directlyrecorded in typical datasets of MNCs. My approach is to use empirical patterns from a firm-leveldataset to inform and discipline a model, with which I will infer indirectly the use of R&D andconduct counterfactual experiments.

Central to my analysis is a newly assembled dataset on R&D and production of MNCs froma broad set of countries. Production data and the ownership network connecting affiliates totheir parents are from the Orbis Database, which has been used in recent studies of MNCs (e.g.,Alviarez, 2019; Cravino and Levchenko, 2017). I link firms in this database to the PATSTATDatabase, a comprehensive database that contains bibliographic information of patents from morethan 90 patent offices. The location of inventors collected at the patent application stage by patentoffices identifies where the R&D underlying each patent is conducted. Aggregating across allpatents granted to a parent firm, I obtain a measure of R&D at the parent-host country level.This measure is robust to how an MNC assign patents to different affiliates within the firm. It isstrongly correlated with bilateral offshore R&D measured using expenditures, but has the benefitof being widely available at the firm level for a broad set of countries.

1DuPont offers a good example. Headquartered in Delaware, U.S., it has major R&D centers located in the U.S.,Brazil, China, Switzerland, Korea, Germany, and Japan. Moreover, it has production facilities in 19 countries, fromwhich it serves around 90 countries.

2

Figure 1: The Level and Growth of Offshore R&D, 1985-2012

Notes: The measure for offshore R&D in country i is Business enterprise R&D expenditures in country i by foreign firmsTotal business enterprise R&D expenditures in country i . Uncolored bars indicate the

value of this variable in 2012; colored bars indicate the value at the beginning of the sample, which differs by country and dates backto as early as 1985. Data source: OECD.

I document three facts using the dataset. First, R&D and production within a firm tend tocolocate. This fact holds both in the cross section and over time. Second, R&D and productionof overseas affiliates decrease in common measures of distance to the headquarters, but the im-portance of individual distance metrics, such as whether a host shares a language with the head-quarters, differs between the two activities. Third, the R&D intensity of an affiliate, measuredas the ratio between R&D and sales, is higher in host countries with better human capital and—within an affiliate—increases as the human capital in the host improves over time. Together, thesefacts highlight three ingredients as necessary for the model: 1) frictions in separating productionfrom R&D, 2) significant but potentially heterogeneous headquarter effects on affiliate R&D andproduction, and 3) the role of talent in R&D.

I incorporate these ingredients into a general equilibrium model of MNCs, which I will use tointerpret the data and to conduct counterfactual experiments. In the model, firms differ along twodimensions of their knowhow: innovation management efficiency—which governs how effectivea firm is in converting researcher input into new product blueprints—and production manage-ment efficiency, which governs a firm’s productivity in converting production labor into output.Firms can enter foreign countries (hosts) to perform offshore R&D, a process that hires researchersin the host as the input and develops new differentiated varieties as the output. Workers differ inability and can choose according to their ability to be a researcher or a production worker.

I embed this offshore R&D decision into a multi-country general equilibrium model of globalproduction and trade (Ramondo and Rodríguez-Clare, 2013; Arkolakis, Ramondo, Rodríguez-Clare and Yeaple, 2018). Specifically, after a product is developed by an R&D center, whetheronshore or offshore, the firm chooses which countries to market it to and for each destination,where production should take place. An American company can thus develop a new product inthe U.K., produce it in China, and export from there to India. The headquarter services embedded

3

in the provision of firm knowhow, R&D, manufacturing production, and marketing constitutesfour stages of a global value chain (GVC) that connects workers and firms from different countries.

The model captures two motives of offshore R&D commonly cited by firms: ‘talent-acquisition’and ‘market-access.’2 The former motive is a response to a host country’s relative abundance ofable researchers, which depends on the availability of researchers and the presence of efficientfirms competing for them. The latter motive reflects the geographic frictions in trade and offshoreproduction: to save on trade and production costs, firms tend to produce near major markets andin countries where wages for production workers are low. The cost of separating production fromR&D in turn gives them an incentive to conduct R&D in these countries.

I calibrate the model to data from 37 major countries. Two crucial fundamentals of the modelare distributions of worker ability and firm knowhow, and the geographic frictions driving firms’R&D and production decisions. I parameterize each country’s distribution of firm knowhow us-ing the World Management Survey built by Bloom et al. (2012) and its talent distribution usinga international cognitive test score database developed by Hanushek and Woessmann (2012). Ispecify the frictions firms face in conducting offshore R&D and offshore production as parametricfunctions of common distance measures. My specification incorporates colocation and the head-quarter effects parsimoniously by parameterizing the friction impeding production of a variety ina host to be a weighted function of that host’s proximity to the headquarters and its proximity tothe R&D center inventing that variety. As the weight varies, the local production share of offshoreR&D ranges from zero to one. I infer the weight and the distance elasticities in these specifica-tions indirectly by matching the geographic coefficients estimated in reduced-form regressions. Idetermine remaining parameters by matching various statistics of the firm size distribution in theU.S. and the overall openness of individual countries to foreign trade, offshore production, andoffshore R&D. The model is able to match untargetd moments closely.

Interpreting the reduced-form patterns through the model, I find that although proximity toboth the headquarters and R&D centers matter for production, the weight on the latter is fourtimes the weight on the former. With this estimate, the model infers on average 70% of R&D inoverseas affiliates is for local production. This share varies across hosts according to its cost ofproduction. In particular, R&D of MNCs in high-income hosts are more likely to lead to produc-tion elsewhere. The model also reveals heterogeneous effects of offshore R&D on the sources ofincome within a country—in the U.S., for example, the profit from products developed at offshorelocations account for about a third of the total profit. In contrast, for firms from most emergingeconomies, this share is zero. As I show later, these patterns, not directly measurable in typicaldatasets, play a vital role for welfare analysis.

I use the model for counterfactual experiments and obtain two sets of quantitative results.The first focuses on the empirical relevance of the two motives of offshore R&D. I weaken the‘talent-acquisition motive’ due to the relative talent abundance of a host by changing its distri-butions of firm knowhow and worker ability. For a typical host country, increasing its average

2According to firm-level surveys (see, e.g., Thursby and Thursby, 2006), access to qualified researchers and hostcountry market potential are the two most important factors firms consider when choosing where to build their offshoreR&D centers.

4

firm knowhow by the interquartile range of the mean knowhow of sample countries decreasesits inward offshore R&D by half; decreases its average worker ability by the interquartile rangeof the mean worker ability reduces inward offshore R&D by two-fifth. Differences in knowhowand ability are thus an important driving force for offshore R&D. I also examine how a country’saccess to foreign consumers through exporting and to foreign producers through offshore produc-tion affect its attractiveness as a destination for R&D. While both consumer and producer accessincrease the return to innovation in a partial equilibrium, they have opposite general equilibriumeffects: losing access to consumers pushes a country to specialize in innovation and to outsourceproduction, whereas losing access to foreign producers pushes it to specialize in production. As aresult, the former increases inward offshore R&D to a host and the latter decreases it. For a typi-cal host, removing access to foreign consumers and producers at the same time decreases inwardoffshore R&D by around 55%. Together, these experiments suggest that the two motives in themodel can account for a significant share of the observed offshore R&D.

The second set of counterfactual experiments focus on normative implications of offshoreR&D. Analytically and quantitatively, I show that an option to conduct R&D abroad generatessignificant welfare gains. The average welfare gains from offshore R&D, defined analogously tothe gains from trade, are around 3.3%. Compared to a restricted model with only trade and off-shore production, incorporating offshore R&D amplifies the gains from openness by a factor of1.3. This amplification is substantially larger for advanced countries than for emerging countries,primarily because a higher fraction of income of advanced countries is from profit and researchersalary generated through offshore R&D. Overlooking this channel will thus not only underesti-mate the gains from globalization, but also bias comparison of the gains across countries.

The relative importance for offshore production of proximity to headquarters and proximityto R&D centers has a significant impact on both the overall gains and its distribution among coun-tries. My estimate suggests that proximity to R&D center matters more; if I had assumed thatonly proximity to the headquarters matter, I would have found much large gains from R&D andopenness, especially for emerging countries. The reason is that under these alternative geographicfrictions, most of offshore R&D will be devoted for production at the headquarters, which in turnaffects the calibrated openness and the distribution of profit across countries. This result stressesthe importance of disciplining the model using micro data.

Existing quantitative studies on multinational activities do not separately model offshore R&Dand offshore production, even though they are very different activities that can be targeted byspecific policies.3 Is this an innocuous assumption for policy simulations? To answer this ques-tion, I consider two experiments that encourage further FDI integration among a set of emergingcountries. The first reduces bilateral offshore R&D costs between these countries and the secondreduce bilateral offshore production costs instead. I find that in the first experiment, the focalcountries benefit at expense of developed countries; in the second experiment, on the other hand,developed countries also benefit significantly. There are losses for developed countries in the first

3For example, countries can grant tax credits or open their borders specifically to R&D intensive FDI. An example isthe U.K. ‘patent box,’ which reduced the corporate tax rate on revenues from R&D by 10 p.p.

5

experiment because in response to the policy, firms from emerging countries conduct more R&D inother emerging countries. This pushes up wages and reduces the profit made from offshore R&Dand production in these countries by firms from developed countries. The second experiment,in contract, makes conducting R&D in emerging countries more profitable for firms from devel-oped countries, as they can now outsource production to nearby emerging countries more easily.This comparison shows policies on offshore R&D and production can have very different effectson third countries. Such differences are especially relevant for studying multilateral investmentagreements.

In addition to the interactions through the general equilibrium, offshore R&D also affects eval-uation of policies on offshore production and trade simply because they are all connected in aGVC. To make this point, I conduct the same bilateral liberalization in offshore production as inthe second experiment before, but in a restricted model without offshore R&D. Different frombefore, this experiment generates welfare losses for developed countries. Integration via offshoreproduction still increases the return to products developed in these countries, but developed coun-try firms do not enjoy direct benefits if they cannot open R&D centers there. By a similar logic,R&D subsidies in a host can have a global influence simply because foreign owners of domesticR&D centers also benefit. These results highlight the importance of incorporating offshore R&D,even if one has no interest in offshore R&D policy per se.

This paper contributes to a growing literature on the impacts of multinational firms (see,among others, McGrattan and Prescott, 2009; Burstein and Monge-Naranjo, 2009; Garetto, 2013;Ramondo and Rodríguez-Clare, 2013; Keller and Yeaple, 2013; Arkolakis, Ramondo, Rodríguez-Clare and Yeaple, 2018; Tintelnot, 2016; Alviarez, 2019). Within this literature, the most closely re-lated papers are Ramondo and Rodríguez-Clare (2013) and Arkolakis, Ramondo, Rodríguez-Clareand Yeaple (2018), both of which study the welfare gains from trade and offshore production. Thepresent paper differs in two aspects. First, I allow firms to mobilize their knowhow abroad towork with foreign researchers. This channel, combined with my parameterization of a GVC thatlinks offshore R&D to production, generates new implications for both the gains from opennessand the effects of specific policy changes. Second, different from Arkolakis et al. (2018) whichtreats the innovation efficiency of a country as a single exogenous parameter, I decompose it intotwo measurable components and examine the role of each in shaping a country’s comparativeadvantage. Also related is Bilir and Morales (2020), which estimates carefully the productivityeffects of R&D within MNCs using data from the U.S. My finding that most offshore R&D is forproduction in the same host is consistent with their result that R&D in an affiliate has a limitedinfluence on other affiliates. Relative to Bilir and Morales (2020), the main difference of my paperis that it formally models firms’ R&D and production decisions and moreover, embeds them ina multi-country general equilibrium setting. These features allow me to answer a broader set ofquestions including the welfare implications of offshore R&D.

This paper also contributes to the literature explaining the pattern of FDI, dating at least asfar back to as the theoretical work by Helpman (1984) and Markusen (1984). More recently, re-searchers have examined the determinants of M&A FDI (e.g., Nocke and Yeaple, 2008), the role of

6

firm heterogeneity (Helpman, Melitz and Yeaple, 2004), the dynamics of MNCs (Garetto, Olden-ski and Ramondo, 2019), and have looked into the role of MNCs in shaping consumer preferenceand creating market access (Head and Mayer, 2019; Wang, 2019).4 The contribution of this paperis two folds. Empirically, I document facts on the joint allocation of R&D and production across abroad set of countries, complementing existing studies, most of which focus on either productionor R&D alone, often using data from a single home or host country.5 Quantitatively, I interpret thepatterns using a tractable model. The model captures complex organizations frequently seen inmodern MNCs and helps me disentangle the importance of being close to the headquarters fromthat of being close to R&D centers, a distinction shown to have critical welfare implications.

Finally, this paper is related to the literature on the measurement of the GVC (see, e.g., Johnsonand Noguera, 2012; Antràs, Chor, Fally and Hillberry, 2012; Koopman, Wang and Wei, 2014; Tim-mer, Erumban, Los, Stehrer and De Vries, 2014; Johnson, 2018). Most existing measurements arebased on sectoral production. Increasingly—at least according to the popular press—countries inthe GVC specialize according to tasks (innovation, production, marketing, etc.) rather than sec-tors.6 One constraint to systematic measurements on country specialization by task is data avail-ability. Indeed, compared to merchandise trade and sectoral production, comprehensive data oncross-border flows of headquarter services and R&D are far more scarce. This paper differs fromexisting work in that, by combining a new dataset on MNC activities with a structural model, itcharacterizes a GVC with its stages differentiated by tasks.7

2 Data and Facts

2.1 Data Sources and Empirical Sample

I assemble a dataset on R&D and production of firms from different countries, linked togetherthrough a firm ownership network. This subsection outlines the data sources and necessarypreparation procedures; Online Appendix A provides details on these procedures and sourcesof supplementary data.

Financial and ownership data. The financial and ownership data are from the Historic Disk ofthe Orbis Database, extracted in April, 2017. I use the 2016 vintage of shareholder data to identifythe parent of each firm, defined as that holding more than 50% control over the firm either directlyor indirectly. These ownership data span across country borders, allowing me to link firms to theirparents whether they are in the same country or overseas. Firms not linked to an identity underthis definition are assumed to be independent.

4See Antràs and Yeaple (2014) for a recent review of the literature on the activities by MNCs.5For example, Irarrazabal, Moxnes and Opromolla (2013) and Keller and Yeaple (2013) document gravity for affiliate

production. On R&D, Branstetter, Li and Veloso (2013) and Kerr and Kerr (2018) focus on international cooperationeither among inventors from different countries, or between inventors and firms from different countries. Both formsof cooperation is often within the boundary of multinational firms.

6The founder of Acer Inc. famously postulated the smile curve, according to which the physical production in themiddle of a value chain is of low value-added while innovation and marketing at the two ends of the chain are of highvalue-added. This phrase quickly caught on among policymakers and popular press.

7This spirit of combining firm-level data and GE models for measurements is related to de Gortari (2019).

7

My primary measure of production is sales, also extracted from the Orbis Historic Database.For robustness I use value added, which has more limited availability. The original financial datacome at year-firm identifier level. I group all firms in a country owned by the same parent intoone and treat this as an affiliate. This step gives me a total of around 185 million parent firm-hostcountry-year observations between 1996 and 2016. The overwhelming majority of parent firmshave only one host—their home countries.

R&D data. I measure firm-level R&D by the output of R&D, granted patents.8 I build thedataset using micro data from the PATSTAT Global Database, which covers patent applicationsacross 90 national and regional patent offices from as early as the 19th century. I match individ-ual patents to firms using a crosswalk from the Orbis Intellectual Property Right Database. Thiscrosswalk links patent applications to firms using a string matching algorithm based on the name,address, and industry classification of the owner. Around 25 million granted patents are matchedto 377,617 unique parent firms in the Orbis Database.

Using patents as a measures for R&D allocation within an MNC has two drawbacks. First,MNCs are in principle free to assign the ownership of patents to any its affiliates, so the locationof the assignee (the owner of a patent) are not necessarily where the R&D was conducted. I usethe location of the inventors instead to circumvent this problem.9 Second, because a patent onlygrants protection within the country in which it is issued, firms often apply for multiple patentsin different countries for the same underlying invention. To avoid double counting, I keep onlyone patent among each patent family covering the same invention.10

I aggregate individual patents to obtain the total number of patents at the level of parent firm,year, and host country, with host countries identified by the location of inventors. In doing so,it is necessary to take a stand on how to classify patents with inventors from different countries.In the baseline analysis, I divide patents based on the number of inventors. For example, if apatent is invented by two inventors in India and one in the U.S., I attribute two-thirds of it toIndia and one-third to the U.S. For robustness, I also perform analyses where I count each of thehosts co-inventing a patent has having invented the entire patent.

Sample country and time. I supplement these firm-level datasets with country characteristics

8The literature has measured R&D activities using both the input (R&D expenditures) and the output (patents). Ichoose patents mainly for its wider availability—whereas firm-level data for R&D of MNC affiliates with more thanone host or one home country are difficult to compile, the information on the universe of patents in most countries ispublicly accessible. In addition, there are two added benefits to patents as a measure. First, MNCs have an incentiveto manipulate R&D expenditures to shift profits across tax jurisdictions. Their incentive to manipulate the locationinformation of inventors appearing on patent application forms, if any, seems weaker as it does not affect the taxes theypay. Second, patent data from different countries are conceptually closer to each other than R&D data. Indeed, R&Din a country trying to catch up with the frontier might be more about learning what others already know than aboutpushing the frontier. Patents in that country, on the other hand, requires pushing the frontier, even if only by a tinystep. In this sense, it is more comparable. Despite the differences between expenditures and patents, existing literatureshows that they are highly correlated at the firm level. Appendix Figure A.1 shows further that for aggregate andbilateral offshore R&D, the two measures are also strongly correlated. The bottom line is even if one prefer to measureoffshore R&D by expenditure, my measure based on patents is a good proxy.

9The location of inventor is from reported address—work or residential—on patent application and is not based onthe nationality of inventors. Assignees and inventors can be in different countries if, for example, an MNC assign allpatents to the headquarters regardless of where they are invented.

10Patents covering the same invention are connected through a common priority, established when the invention wasfirst filed for patent protection in any country.

8

Table 1: Sample Structure: Firms, Production Facilities, and R&D Centers

(1) (2) (3) (4)R&D center count

Period Firm count Prod. facility count Baseline Liberal1996-2000 26,635 43,395 30,506 36,8182001-2005 48,473 76,641 52,243 61,5392006-2010 75,336 113,401 78,924 91,7032011-2016 86,335 131,493 88,569 102,886Total 236,779 364,930 250,242 292,946

Note: Each row is for a five-year interval between 1996 to 2016. The table reports the numbers of distinct firms (Column1), distinct production facilities (Column 2), and distinct R&D centers (Columns 3 and 4) in each period in the R&Dand financial matched sample. Under the baseline definition, a country qualifies as a host of offshore R&D if the firminvents at least a complete patent in that country; under the liberal definition a country qualifies if it has made a partialcontribution to a patent—e.g., through working with inventors residing in other countries.

and various measures of geographic frictions between countries. To have a consistent sample asthe quantitative section, I restrict to 37 host countries selected based on the coverage of the WorldInput-Output Database and the Penn World Table, both of which are crucial for calibration (seeTable 9 for the country list). The firm-level data cover a substantial share of the economy of thesecountries: in all but three (the U.S., Turkey, and Canada), the total sales of firms in the OrbisDatabase in 2016 are well above 50% of GDP. To reduce measurement errors from my discretemeasure of R&D and to smooth out yearly fluctuations in country characteristics, I aggregate datato four five-year intervals between 1996 and 2016.

2.2 Descriptive Statistics

I combine the financial and patent data. The majority of firms never patented; some firms do-ing R&D have no available financial information. For most of the following empirical analysis, Ifocus only on firms that both have some financial information and have filed at least one patent—invented any where—during the sample period; for tabulating country-level statistics in quantifi-cation later, I will use the full set of firms to ensure good representation of the economy. The restof this subsection provides descriptive statistics on the combined R&D and production sample. Inthe appendix, I provide similar tabulations separately for the full set of firms.

Table 1 is an account of sample coverage. The first column is the number of parent firms ineach period. It increases from around twenty-six thousands during 1996-2000 to around eight-sixthousand in 2011-2016, reflecting the broadening coverage of the database. The second columnis the number of production affiliates. I define a host as a production affiliate of a firm if it haspositive sales over the period. On average, each firm operates in 1.5 countries. The third column isthe R&D center count. As a baseline, I define a host country as an R&D center for a firm if it inventsat least one full patent there. According to this definition, firms on average has 1.05 R&D centers.Columns 4 provides R&D center counts defined more liberally, under which a host country is anR&D center as long as it invents a positive fraction of a patent. This definition includes hostswhose patent count is positive but smaller than one, which can happen if, for example, it invents asmall number of patents with co-inventors from headquarters. Under this definition, the average

9

Table 2: Firms by Number of Affiliates

(1) (2) (3)Firms with Firms with # R&D centers

# # production site baseline liberal0 2,195 18,430 2,6301 207,657 205,539 210,7022 9,910 7,321 13,7713 4,119 2,049 3,7874 2,465 1,017 1,7445 1,672 660 1,055≥6 8,761 1,763 3,090Total 236,779 236,779 236,779

Note: This table counts the number of firms with different numbers of production affiliates and R&D centers. Forexample, 207,657 firm-period observations has exactly one production site. The tabulation pools al four periods, so thesame firms from different periods are treated as two observations.

R&D center for per firm is 1.4.Table 2 tabulates firms with different numbers of production affiliates and R&D centers. The

first row is the number of firms with no production affiliates or R&D center in the 37 countries.11

The second row indicates the number of firms with exactly one production facility or one R&Dcenter. Most firms fall into this category. About 5% of firms have production or R&D activities inmore than one country. Production is generally more spread out than R&D. Around 4% of firmsproduce in more than 6 countries but only less than 1% of firms conduct R&D (under the baselinemeasure) in more than 6 countries.

2.3 Three Facts on Location Choices of R&D and Production

I document three facts on the spatial distribution of R&D and production within a MNC.Fact 1: R&D and production tend to be located together. The first fact is on the co-location of

R&D and production. The specification establishing this fact takes the following form:

y f h,t = FE + γR&D x f h,t +−→γ dist ·

−→dist f h,t + ε f h,t,

where f , h, t, indicates firm, host, and period, respectively. FE are a set of fixed effects. Variablesy f h,t and x f h,t measure production and R&D of firm f in host h.

−→dist f h,t is a vector consisting of

measures of the average distance between host h and all other countries i in which firm f has anR&D center. Four distance measures are included: the greater circle distance, and indicators forwhether country h and the R&D center country share an official language, are contiguous, or havea colonial tie.

−→dist f h,t is firm specific because firms differ in the geographic distribution of R&D

centers.Coefficients γR&D and −→γ dist capture the colocation patterns. To rule out the influence on these

coefficients of country or industry characteristics, all specifications include host-period, home-

11All firms in the sample have at least one production affiliate and one R&D center (under the liberal definition), butthese might not be in the 37 countries considered here.

10

Table 3: Cost of Separating R&D from Production

(1) (2) (3) (4) (5) (6)Dependent var. prod. indicator ln (sales)

R&D Indicator f h,t 0.281∗∗∗ 1.165∗∗∗ 1.042∗∗∗

(0.003) (0.023) (0.026)ln(patent) f h,t 0.333∗∗∗ 0.330∗∗∗ 0.205∗∗∗

(0.012) (0.012) (0.044)ln (distance) f h,t -0.023 -0.339∗∗

(0.025) (0.144)common language f h,t 0.143∗∗∗ 0.184

(0.049) (0.237)contiguity f h,t 0.220∗∗∗ 0.376

(0.051) (0.263)colonial tie f h,t 0.091∗∗ -0.493

(0.046) (0.305)R2 0.704 0.495 0.496 0.571 0.571 0.969Within R2 0.042 0.045 0.047 0.093 0.094 0.022Observations 7494794 119736 119592 19548 19548 14090Firm-period FE Y Y Y Y Y YHost-period FE Y Y Y Y Y YHome-host FE Y Y Y Y YHost-industry FE Y Y Y Y YAffiliate FE Y

Note: Columns 1 estimates the relationship between having an R&D center in a host and the probability of having aproduction facility in the same host. Columns 2 through 6 estimate the relationship between having an R&D center(and the size of the R&D center) and the size of the production facility in the same host. Columns 3 and 5 also controlfor average distance of a production facility to other countries in which the firm has an R&D center.Standard errors (clustered at firm level) in parenthesis. ∗ p < 0.10, ∗∗ p < 0.05, ∗∗∗ p < 0.01.

host, and host-industry fixed effects. Further, I control for firm-period fixed effects, so the varia-tions exploited are within a firm, across host countries.

Table 3 reports the results. The first column focuses on the extensive margin and shows thathaving an R&D center in a host increase the probability of having a production affiliate in the samecountry by 0.28, or about ten times the mean value of the outcome variable. Column 2 focuses onthe relationship between having an R&D center and the production of an affiliate in the samehost—host countries with an R&D center on average produce 116% more. Column 3 includes theaverage value of the four distance measures between h and all countries in which firm f has anR&D center. The R&D indicator coefficient shrinks slightly. The distance coefficients show thatbeyond the proximity to the R&D center in the same host, being close to sibling R&D centers inother countries also have an effect. Columns 4 and 5 use log patent count as the explanatoryvariable. The sample size is smaller as only affiliates with positive R&D are included, but themain point stands. Having a larger R&D center in the same host and being close to the siblingR&D centers are both correlated with affiliate production.

The results so far are consistent with existence of frictions impeding separation of R&D andproduction, but they can also be driven by idiosyncratic match quality between a firm and a hostbeyond what can be controlled through industry fixed effects. For example, technology, humancapital, or infrastructures of a country might be particularly conducive for some firms because

11

they fit the technology of the firm, which benefits both R&D and production. To alleviate thisconcern, Column 6 controls for affiliate fixed effects, so it exploits the correlation in over-timechanges of production and R&D in a host, controlling for the overall trend of the host and theindustry, and the growth of the firm common to all its affiliates. The estimated elasticity is around17%, so affiliates doing more R&D also see a larger increase in production.12

Given the colocation of R&D and production, it might be tempting to think that it is withoutloss of generality to bundle R&D and production as a composite ‘production’ activity and modelsMNCs as choosing where to locate this composite activity. Facts 2 and 3 show that, in contrast towhat such a model would predict, R&D and production of overseas affiliates respond differentlyto geographic frictions and other types of country characteristics.

Fact 2: Affiliate production and R&D decrease with distance to the headquarters, but atdifferent rates. The second fact concerns the headquarter effects on affiliate R&D and production.The specification is the following:

y f oh,t = FE +−→γ dist ·−→distoh + εoh f ,t,

in which the dependent variable y f oh,t is measures of activities in host h of firm f from countryo in period t.

−→distoh, a vector of four distance measures between headquarters o and host h, is

the focus of this specification. Different from in Fact 1, the distance measures here vary onlyacross country pairs but not firms. I control for firm-period, host-industry, and host-period fixedeffects throughout. Further, I exclude headquarters from the sample, so the comparison is amongaffiliates of the same firm in different countries.

Columns 1 and 3 of Table 4 report the results of a linear probability model in which the de-pendent variable is an indicator for having a production affiliate or an R&D center in h. Theregressions suggest geographic frictions play important but heterogeneous roles. Sharing a lan-guage is important for both R&D and production, whereas distance and colonial ties matter morefor production. The mean value of the dependent variable is 0.027 and 0.018, respectively, so theestimated coefficient are economically sizable. Columns 2 and 5 estimate the intensive margineffects on affiliate R&D and production of distance to the headquarters. According to the esti-mates, having a common language is more important for R&D, but other measures of geographicfrictions are in general more important for production than for R&D.

These results are not sufficient to conclude that distance to the headquarters matters for bothR&D and production—given Fact 1, as long as one activity is hindered by distance to the head-quarters, regressions on the other activity will show expected signs for geographic frictions. Toshow qualitatively both headquarters effects are at play, Columns 4 and 6 add the indicator forR&D center for extensive and intensive margin regression respectively. Consistent with coloca-

12Two remarks on this specification. First, when exploiting over-time variations firms grow from a small numberof patents can have an extraordinary percentage growth rate. To avoid these firms having an out-sized impact on theestimate, I weight firms by the square root of their patent counts. Not weighting would result in an estimate of 0.06 witha t statistics of 3.92. Second, I do not include the average distance metrics as there are not enough overtime variationsto estimate them precisely. The inclusion of these variables, however, does not impact the estimated coefficient forlog(patent) materially.

12

Table 4: The Headquarter Effect on R&D and Production

(1) (2) (3) (4) (5) (6)R&D Production

R&D indicator ln(patent) prod. indicator ln (sales)

ln(distance)oh -0.002∗∗ -0.129∗∗∗ -0.005∗∗∗ -0.005∗∗∗ -0.284∗∗∗ -0.254∗∗∗

(0.001) (0.035) (0.002) (0.001) (0.028) (0.027)contiguityoh 0.002 0.104 0.004 0.003 0.183∗∗∗ 0.172∗∗∗

(0.002) (0.073) (0.004) (0.004) (0.064) (0.058)common languageoh 0.020∗∗∗ 0.266∗∗∗ 0.022∗∗∗ 0.015∗∗ 0.159∗∗ 0.091

(0.004) (0.073) (0.009) (0.007) (0.068) (0.061)colonial tieoh 0.002 0.020 0.024∗∗∗ 0.023∗∗∗ 0.152∗ 0.128∗

(0.004) (0.067) (0.008) (0.007) (0.079) (0.068)R&D indicator f h,t 0.375∗∗∗ 1.198∗∗∗

(0.019) (0.031)Observations 7294921 45386 7294921 7294921 103198 103198R2 0.124 0.336 0.303 0.339 0.420 0.445Within R2 0.004 0.012 0.006 0.058 0.012 0.056Firm-period FE Y Y Y Y Y YHost-industry FE Y Y Y Y Y YHost-period FE Y Y Y Y Y Y

Note: Columns 1 and 2 estimate the relationship between the proximity of a host to the headquarters and affiliate R&D.Columns 3 to 6 estimate the relationship between the proximity to the headquarters and affiliate production, amongwhich Columns 4 and 6 also include the R&D center indicator. Observations of the headquarters of firms are excludedfrom all regressions.Standard errors (clustered at country-pair level) in parenthesis. ∗ p < 0.10, ∗∗ p < 0.05, ∗∗∗ p < 0.01.

tion effect, the estimates for R&D indicator is statistically significant and similar to Table 3. Theestimated geographic coefficients decrease, which suggest indeed some of the results are drivenby colocation. Yet the coefficients remain statistically significant, consistent with a headquarter ef-fect for production. More generally, without additional assumptions it is difficult to quantitativelydisentangle colocation from the headquarter effects for production and R&D. I will unpack theseforces in quantification by combining these estimates with a structural model.

Fact 3: The R&D intensity of an affiliate increases in host country human capital quality.The third fact looks into how the R&D intensity of an affiliate varies with host characteristics. Ipay particular attention to host human capital, as when interviewed about the determinants ofthe location choice of R&D centers, mangers of large multinational firms frequently rate access totalent as a primary factor (Thursby and Thursby, 2006). My specification is as follows:

y f oht,t = FE +−→γx ·−→X oh,t + εoh f ,t.

The dependent variable is the log ratio of affiliate R&D over sales. The explanatory variables−→X oh,t

include the measure of host talent, the human capital index from the Penn World Table, along withother controls.

Table 5 reports the results. The first column controls for firm-period fixed effects and various

13

Table 5: Human Capital and Affiliate R&D Intensity

(1) (2) (3) (4)Dependent variable: ln (patent/sales)human capital 0.979∗∗∗ 2.598∗ 3.013∗∗ 3.431∗∗

(0.212) (1.423) (1.375) (1.353)ln(GDP per capita) -0.308 -0.188 -0.622 -0.710∗∗

(0.217) (0.501) (0.453) (0.352)IPR protection 0.563∗∗ 0.404∗

(0.235) (0.220)R&D subsidies 0.508 0.572

(0.446) (0.465)ln (researchers) 0.421∗∗

(0.211)ln (GDP) 0.057

(0.087)Observations 21031 14100 11803 11464R2 0.252 0.715 0.753 0.754Within R2 0.029 0.005 0.010 0.015Distance measures Y - - -Firm-period FE Y Y Y YAffiliate FE - Y Y Y

Note: The dependent variable is log of the ratio between patent counts and affiliate sales and the independent variablesare country characteristics. Column 1 is a cross-sectional regression that controls for firm-period fixed effects andbilateral distance measures (See Table 4 for these measures). Columns 2 through 4 control for time-invariant hostcharacteristics through affiliate fixed effects. Because over-time variations in GDP and GDP per capita are highlycorrelated, Columns 2 to 4 only includes the latter.Standard errors (clustered at host-country level) in parenthesis. ∗ p < 0.10, ∗∗ p < 0.05, ∗∗∗ p < 0.01.

measures of bilateral distance. The estimate indicates a significant positive correlation betweenR&D intensity of an affiliate and the human capital index. The size and income of the host, onthe other hand, does not seems to be important. Columns 2 adds affiliate fixed effects and showsthat as the human capital of a host improves, affiliates there become more specialized in R&D.13

Column 3 further controls for protection of intellectual property rights (Park, 2008) and R&D sub-sidies (OECD) in a country, two policy measures likely correlated with R&D. The coefficient forIPR protection is statistically significant, but it does not affect the coefficient of human capital.Finally, Column 4 further includes a narrower measure of talent, the number of researchers in acountry. An increase in total number of researchers in country is positively correlated with theR&D intensity of affiliates there, but the coefficient for human capital index scarcely changes, sug-gesting that a broader interpretation of talent is warranted.

Robustness and summary. In the appendix, I show that the three facts are robust to differentmeasures of affiliate R&D and production, and alternative specifications. Together, these factsunderscore two motives for firms to conduct R&D in a host: the incentive to be close to productionand the incentive to take advantage of host talent. Moreover, the differential headquarter effects

13This regression exploits over-time variations in human capital of a host. Because changes in GDP and income arehighly correlated, I include only one of them. Including both does not affect the point estimate for human capital andits statistical significance materially.

14

and the variations in R&D intensity across countries shows that while R&D and production arecertainty interconnected decisions, firms do have the flexibility in choosing the appropriate sitefor each separately.

The facts call for a model in which R&D and production are explicitly introduced as separatedecisions that respond to different incentives. In the rest of this paper, I develop such a model tointerpret the data and to quantify the aggregate implications of offshore R&D.

3 The Model

3.1 Environment

Worker. There are N countries, indexed by i = 1, 2, ...N. Country i is endowed with Li measureof workers, whose ability α is from a distribution with a cumulative distribution function (CDF)denoted by Ai(α).14 Workers with ability α choose weather to work in manufacturing production,where everyone has the same effective productivity and earns a common wage of W l

i (l for low-skill) independent of ability, or to work on more skill-intensive, ‘high value-added’ tasks—R&Dand marketing—and earn a wage of Wh

i × α. Letting αi denotes the ability level above which aworker chooses high-skill occupations, we have

Whi αi = W l

i

The effective supply of high and low skill efficiency units are denoted by Lhi and Ll

i , respectively,and given by:

Lhi = Li ·

∫α>αi

α dAi(α)

Lli = Li · Ai(αi).

Firm. Country i is endowed with Ei measure of firms differing in their innovation managementefficiency, zR ∈ ZR, with a CDF denoted by GE

i (zR). GE

i (zR) captures the state of knowhow in

country i, which I interpret as adopted management practice that is conductive for innovation,but can also be thought of more broadly as technological expertise accumulated within a firm.Firms build R&D centers in different countries, which recruit local researchers to develop newvarieties, and then engage in offshore production and trade.

Consumption. Workers decide how to allocate their income according the preference repre-sented by the following utility function :

Ui = (∫

Ωi

qi(ω)σ−1

σ dω)σ

σ−1 ,

14The talent distribution in a country reflects the quality of the education system, education choice, as well as culturaltraits such as openness to innovation. By taking the talent distribution as given, this paper abstracts from the effect ofinternational integration on these factors.

15

where Ωi denotes the set of product varieties available in country i, qi(ω) is the consumption ofvariety ω, and σ > 1 is the elasticity of substitution. Let the aggregate consumption expenditurein country i be Xi. The demand for variety ω is:

qi(ω) = pi(ω)−σ Xi

Pi1−σ

,

with Pi1−σ =

∫Ωi

pi(ω)1−σdω being the ideal demand price index aggregated over pi(ω), the priceof variety ω in country i.

3.2 Firm Decisions: Overview

Firms operate in multiple countries and make decisions on R&D, production, and exporting. Thissubsection gives an overview of these decisions. I will use the following indexing conventionsthroughout: o denotes a firm’s headquarters, that is, the country where a firm originates andobtains its draw of zR; i denotes the country where a product is developed—the location of theR&D center; m denotes the country where manufacturing production takes place; and d denotesthe destination country where a product is consumed.

Consider a firm from country o. Knowing its innovation efficiency in the home country, zR, thefirm decides how many R&D centers to open and in which countries. To open an R&D center incountry i, it pays a fixed cost of cR

oi in country i high-skill efficiency unit. As the subscript indicates,this cost varies across pairs of countries. The innovation management efficiency of this R&D centerdepends on that of its parent.15 Motivated by Fact 2, I assume that firms can only partially transfertheir efficiency to offshore R&D centers. Letting φR

oi ≤ 1 be the proportion of innovation efficiencythat can be transferred, the innovation efficiency for an R&D center in country i operated by acountry o firm is zR = zRφR

oi. This efficiency governs how many varieties can be developed for agiven number of researchers.

Innovative firms are not always the most efficient in carrying out manufacturing. To allow forthis heterogeneity, each R&D center upon entry also obtains a random draw of production man-agement efficiency, denoted zP ∈ ZP, which is common to all products developed by the R&Dcenter. To capture positive correlation between innovation efficiency and production efficiency,the distribution from which zP is drawn increases in zR in the sense of first-order stochastic dom-inance. I use GP(zP|zR) to denote the CDF for production efficiency draws, with gP(zP|zR) beingthe corresponding probability density function (PDF).16

Firms may open multiple R&D centers in different countries, but at most one in each. Giventhe production and innovation efficiency of an R&D center, the parent firm decides how manyresearchers to recruit for R&D and which countries to sell their products to. To sell products todestination country d, a per-variety fixed marketing cost of cM

d in terms of country d production

15This assumption follows earlier work on theory of multinationals, see, e.g., Helpman et al. (2004).16Under this setup, R&D centers belonging to the same parent but with different innovation efficiencies are drawing

zP from different distributions. Production management efficiency is therefore specific to individual R&D centers. Analternative interpretation of this efficiency is the quality of products developed by an R&D center.

16

Figure 2: Overview of the Model

Home Country

o

Host Country

𝑖1

Market 𝑑1

Prod. site 𝑚1

Headquarters R&D Production Market

Host Country

𝑖2

(𝑧𝑃, 𝑧𝑅)

Innovation Efficiency ǁ𝑧𝑅

Draw (𝜂1, 𝜂2, 𝜂3, … , 𝜂𝑁)

Shipping cost 𝜏𝑚1𝑑1

Unit Delivery Cost𝑤𝑚1𝑙 𝜏𝑚1𝑑1

𝑇𝑚1∙𝑧𝑃𝜙𝑜𝑖1𝑚1

𝑃 ∙𝜂𝑚

Marketing

cost 𝑐𝑑1𝑀

labor needs to be paid.Firms can manufacture products in countries where they do not necessarily perform R&D, and

then export to destination countries. By separating production from R&D and the headquarters,firms take advantage of cheaper production labor and save on shipping fees. However, geographicseparation makes it difficult for researchers and mangers to communicate with the production line,reducing production efficiency. I use φP

oim ≤ 1 to denote the fraction of productivity that a firmcan transfer to its production center. In keeping with the empirical facts, I will parameterize φP

oim

to be a function of both the distance between o and m and the distance between i and m, allowingproximity to both the headquarters and to R&D centers to play a role.

Production uses only labor. For an R&D center with production efficiency zP, the deterministiccomponent of the productivity with which its varieties can be produced in country m is Tm · zPφP

oim,with Tm being the country-specific manufacturing efficiency at m. I further assume that there is astochastic element, ηm, idiosyncratic to individual varieties, which enters productivity multiplica-tively, so the variety-level productivity in l is Tm · zPφP

oimηm. The cost of producing one unit in mand delivering it to d is W l

mτmdTmzPφP

oimηm, which takes into account the wage for low-skill worker, W l

m,and iceberg shipping fee, τmd.

The structure of the model is illustrated in Figure 2. Importantly, I assume that different vari-eties developed by a firm, either in the same or in different R&D centers, are differentiated fromeach other and from varieties developed by all other firms. This assumption simplifies the modeland is also consistent with how R&D is organized in many multinational firms. General Elec-tric, for example, organizes its ten research labs by scientific disciplines in five countries (the U.S.,

17

Germany, India, China, and Brazil).17 18 This assumption implies that firms make offshore R&Ddecisions for each country independently and that R&D centers affiliated with the same firm op-erate as if they are independent from each other.

Given this independence, in the remainder of this section, I first explain the production andtrade decisions of a firm after a variety has been developed. I then describe the innovation deci-sion of each R&D center and firms’ decisions to build offshore R&D centers. Finally, I define theequilibrium and characterize the welfare gains from openness.

3.3 Production and Trade

Consider a variety developed by an R&D center (zP, zR) in country i from o. To produce it, theR&D center obtains a vector of N idiosyncratic productivity draws, one for each potential produc-tion site, denoted ηηη = (η1, η2, .., ηN) ∈ HHH. The cost of serving country d by producing in country lis coilm = W l

mτmdTmzPφP

oimηm.

Knowing the realization of ηηη, the firm decides whether to sell a variety to destination market d.Entry to d for each variety requires hiring cM

d high-skill workers from d—with a wage of Whd —for

marketing. After paying this marketing cost, the firm chooses the lowest cost production locationfor each of its varieties. Because there are no fixed costs in offshore production, all countries arepotential production sites. With monopolistic competitive markets, the price of this variety incountry d is a constant markup over the lowest unit delivery cost among all choices:

poid(zP, ηηη) = minmσ

σ− 1W l

mτmd

TmzPφPoimηm

.

The operation profit from manufacturing this variety and selling it to country d is

πoid(zP, ηηη) = 1(poid(zP, ηηη) < pd) · [1σ

Xd

P1−σd

poid(zP, ηηη)1−σ − cMd Wh

d ],

where pd ≡(

σWhd cM

dXd

) 11−σ

Pd is the cutoff price below which a firm can sell enough to recover the

17Alternatively, this assumption can be interpreted as capturing M&A FDI. More than 70% of FDI flows in the dataare in the form of mergers and acquisitions (Nocke and Yeaple, 2008). One rationale for these activities is that acquiringfirms can improve the efficiency of the targets by instilling in managerial knowhow. The differentiated-variety assump-tion adopted in the present paper is consistent with this perspective of FDI—multinationals transfer their knowhow tonewly acquired foreign firms, which have their own brands and products, and help them carry out independent prod-uct development. Empirically, Guadalupe et al. (2012) documents an increase in innovation and adoption of foreigntechnology upon acquisition by foreign companies.

18This assumption treats R&D at headquarters and R&D in offshore centers symmetrically. Bilir and Morales (2020)find that for the U.S. multinationals, R&D at headquarters have stronger spillover effects to their affiliates overseasthan R&D at one affiliate to other affiliates. While the current model cannot account for this finding, an extension ofthe model that allows firms to first invest in R&D to build up ‘core management capacity’ before performing productinnovation at home and abroad would be consistent with it. Whether we treat R&D at the headquarters differently willnot affect the qualitative conclusions from the empirical section as firm fixed effects are incuded in al specifications;my quantitative experiment based on the present model can be viewed as a short-term version that does not allow forendogenous accumulation of the core innovation efficiency.

18

marketing cost. The sales revenue of the variety is

roid(zP, ηηη) = 1(poid(zP, ηηη) < pd) ·Xd

P1−σd

poid(zP, ηηη)1−σ.

Letting πoid(zP), roid(zP), and cMoid(z

P) denote the expected values of operation profit, salesrevenue, and marketing cost associated with market d before ηηη realizes and letting F be the CDFof ηηη, we have:

roid(zP) ≡ E(roid(zP, ηηη)|zP) =Xd

P1−σd

∫HHH1(poid(zP, ηηη) < pd)poid(zP, ηηη)

1−σdF(ηηη) (1)

cMoid(z

P) ≡ E(cMoid(z

P, ηηη)|zP) = cMd wH

d

∫HHH1(poid(zP, ηηη) < pd)dF(ηηη)

πoid(zP) ≡ E(πoid(zP, ηηη)|zP) =1σ

roid(zP)− cMoid(z

P).

For tractable aggregation, I assume the following throughout the rest of this paper:

Assumption 1.

a) F(xxx) ≡ Prob(η1 ≤ x1, ..., ηN ≤ xN) =

1− (

N

∑m=1

1N

x−θm ), ∀m ∈ 1, ...N, xm ≥ 1

0, ∃m ∈ 1, ..., N, xm < 1

b) pd <σ

σ− 11zP

W lmτmd

TmφPoim

, ∀ o, i, m, d, zP ∈ Zp

F(xxx) is a special case of the multivariate distribution proposed in Arkolakis et al. (2018) andhas two attractive implications. First, poid(zP, ηηη)1−σ follows a Pareto distribution. Second, togetherwith Assumption 1.b, it implies that the expected fraction of sales in country d that is fulfilledthrough production in country m, denoted by ψoimd, is19

ψoimd =(

TmφPoim

W lmτmd

)θ

ζθoid

, where ζθoid ≡ [∑

m

1N(

Tmφpoim

W lmτmd

)θ ]1θ . (2)

Common to all production sites, zP does not enter the expression.

19More precisely, poid(zP, ηηη)1−σ follows a Pareto away from its minimum support. Assumption 1.b implies that amongfirms actively selling to d, that minimum support never binds. The economic content of this assumption is that themarketing cost is high enough so that for all firms there are bad enough realization of ηm such that they will not enter d.As long as this assumption is maintained, the exact value of cM

d does not matter for quantitative analysis. For example,cM

d could vary by firm.

19

In the appendix, I show that under Assumption 1, Equation (1) simplifies to:

roid(zP) =θ(σ− 1)θσ1− θσ

σ−1

θ − (σ− 1)X

θσ−1d Pθ

d (wHd cM

d )θ+1−σ

1−σ (ζoidzP)θ (3)

cMoid(z

P) =θ − (σ− 1)

θσroid(zP)

πoid(zP) =1σ

roid(zP)− cMoid(z

P) =σ− 1

θσroid(zP).

The total operation profit from a variety by an R&D center in country i is therefore

πoi(zP) = ∑d

πoid(zP). (4)

3.4 R&D Decisions

R&D centers hire high-skill workers to develop new varieties. Letting h be the measure of high-skill efficiency units recruited, the total number of new varieties developed, v, is:20

v = zR · hγ.

Two comments are in order regarding this setup. First, I model R&D as invention of differentialproducts. By taking this stand, my model speakers to large conglomerates that organize globalR&D by product lines or discipline. An alternative is to introduce R&D as process innovation,which improves manufacturing efficiency of both the affiliate in the same host and affiliates else-where through spillovers. I wish to note that such spillovers are internalized within the firm andare also present in my setup, where the invention of an R&D center influence the sales of affiliatesin other countries because they too can produce these varieties. Ultimately, both models implythat R&D in a host can increase revenues of the firm both locally and globally, and without directmeasures of ‘products,’ it is difficult to distinguish the two. In quantification, I will map the num-ber of varieties to number of patents, which recent evidence has found to be a stronger predictorof new product introduction (Argente et al., 2018).

Second, I assume 0 < γ < 1, implying decreasing return to the size of research teams in anR&D center. γ captures the importance of the fixed factor in R&D from the firm side, embodiedin zR. An example of such fixed factors is the capacity of top management in supervising thedevelopment effort, which does not scale easily.21

20In an earlier version of this paper (Fan, 2017), I assume that the quantity and quality of researchers are imperfectsubstitutes and that there is log supermodularity between zR and researcher ability. That assumption leads to positiveassortative matching between firms and researchers and have additional implications on firms’ location choice. Giventhat most of the qualitative predictions of the model do not depends on that assumption, the present paper abstractsfrom this feature for simplicity.

21See Antras et al. (2006) for an analysis of an offshoring model in which managers can only supervise a fixed numberof workers. Alternatively, decreasing returns to scale might stem from increases in coordination costs, free-riding, anddisagreement among researchers as teams expand. Such coordination costs have been documented empirically. Forexample, Haas and Choudhury (2015) finds that, while total patenting increases with the number of members in ateam, the increase is smaller than the increase in the team size

20

The optimization problem for the R&D center is

πRoi(z

P, zR) = maxhπoi(zP)zRhγ −Whi h,

where πoi(zP) is the expected per-variety profit. The first order condition implies

hoi(zP, zR) =(γπoi(zP)zR

Whi

) 11−γ

,

which gives the number of varieties invented as:

voi(zP, zR) = zR1

1−γ

(γπoi(zP)

Whi

) γ1−γ

. (5)

The total operation profit from these varieties is πoi(zP) · voi(zP, zR). Of this, a γ fraction goes toresearchers. The rest is the return to firms’ knowhow, given by:

πRoi(z

P, zR) = (γγ

1−γ − γ1

1−γ )( 1

Whi

) γ1−γ (

πoi(zP)zR) 11−γ . (6)

This return also represent the total variable profit from doing R&D in country i.

3.5 Offshore R&D and the Distribution of R&D Centers

Let πRoi(z

R) be the expected profit (across all possible zP draws) for an R&D center in country i,given innovation efficiency zR. It is given by:

πRoi(z

R) =∫

ZPπR

oi(zP, zR)gP(zP|zR)dzP

To characterize firms’ decision to open offshore R&D centers, I assume the following about gP(zP|zR).

Assumption 2. The distribution from which an R&D center draws its production efficiency zP increasesin the innovation efficiency of the R&D center in the sense of first-order stochastic dominance.

Firms compare the expected profit from building an offshore R&D center to the fixed cost ofsetting up the center, cR

oiWhi . By Equations (3) and (6) , πR

oi(zP, zR) increases in zP and zR. As-

sumption 2 then implies that πRi (z

R) increases in zR, so there exists a cutoff zRoi, so that firms from

country o will perform offshore R&D in country i if and only if its innovation efficiency at theheadquarters zR is above zR

oi. This cutoff is given by:

πRoi(z

Roiφ

Roi) = cR

oiWhi .

The measure of R&D centers from o to i, denoted by Roi is thus:

Roi = Eo ·(

1− GEo (z

Roi))

,

21

in which Eo is the measure of firms from country o and GEo (zR) is the CDF of the innovation

efficiency of these firms at home. The CDF of the innovation efficiency of these R&D centers,denoted GR

oi, is characterized by:

RoiGRoi(z

R) = 1(zR > zRoiφ

Roi) · Eo · GE

o( zR

φRoi

).

The associated PDF is:

gRoi(z

R) =1

Roi1(zR > zR

oiφRoi) · EogE

o (zR

φRoi)

1φR

oi.

Finally, the density distribution of the R&D centers from country o over (zP, zR) is:

goi(zP, zR) = gP(zP|zR)gRoi(z

R).

3.6 Aggregation

I now derive the expressions for aggregate outcomes. Letting Voi(zP) be the marginal density ofvarieties invented in country i by R&D centers from country o with production efficiency zP, i.e.,

Voi(zP) = Roi

∫ZR

voi(zP, zR) · goi(zP, zR)dzR

where voi(zP, zR) is given by Equation (5). The total measure of varieties across all zP is:

Voi =∫

ZZZPVoi(zP)dzP.

In the appendix, I show that the price index in country d is:

P1−σd = θ(

σ

σ− 1)−θ 1

θ − (σ− 1)

(σWhd cM

dXd

) θ−(σ−1)1−σ

Pdθ−(σ−1) ∑

o∑

iζθ

oid

∫ZZZP(zP)θVoi(zP)dzP. (7)

The sales to country d of the varieties developed in i by firms from o, denoted by Xoid, is

Xoid = θ(σ

σ− 1)−θ 1

θ − (σ− 1)(σWh

d cMd )

θ−(σ−1)1−σ

( Xd

P1−σd

) θσ−1

ζθoid

∫ZZZP(zP)θVoi(zP)dzP. (8)

These sales can be fulfilled through production in any country. Let Xoimd be the value of produc-tion in country l. Because Equation (2) holds for each individual variety making up Xoid, it appliesto the aggregate flows, i.e.,

Xoimd = ψoimdXoid.

The income from these sales are split among different participants along the four-stage value

22

chain. From Equation (4), the amount accrued to each participant is a direct function of Xoimd.First, the bulk of sales ( σ−1

σ ) is the manufacturing value added taking place in country m. LettingYom be the total manufacturing income generated through production of firms from o in countrym, we have

Yom =σ− 1

σ ∑i,d

Xoimd

The markup is split among three participants. First, a θ−(σ−1)θ fraction is used for marketing in

destination d. The total marketing costs incurred in country d by firms from o, denoted by CMod , is:

CMod =

1σ·( θ − (σ− 1)

θ

)∑i,m

Xoimd.

The reminder of markup is value added by researchers in country i—in the forms of both productdevelopment and the overhead of R&D centers—and the net profit to firm owner at o. Denotingthe variable profit of firms from o operating R&D centers in country i by Πoi, we have:

Πoi =1− γ

σ·(σ− 1

θ

)∑m,d

Xoimd.

The payment to researchers directly working on product developed in country, denoted by Ioi, is:

Ioi =γ

1− γΠoi.

The payment to high-skilled workers in the form of overhead R&D cost is

CRoi = Eo · [1− GR

o (zRod)] · cR

oiWhi .

The labor market clearing condition for low-skill workers is

W ldLl

d = ∑o

Yod

The labor market clearing condition for high-skill workers is

Whd Lh

d = ∑o

Iod︸ ︷︷ ︸product development

+ ∑o

CMod︸ ︷︷ ︸

marketing entry

+ ∑o

CRod︸ ︷︷ ︸

RD overhead

.

Total income equals total expenditures

Xd = Whd Lh

d + W ldLl

d︸ ︷︷ ︸Labor Income

+∑i(Πdi − CR

di)︸ ︷︷ ︸Net Profits

Definition 1. The competitive equilibrium is defined as a set of allocations and prices, such that workers

23

and firms optimize, all markets clear, and all firm-level decisions are consistent with aggregate allocationsand prices (see appendix for details).

3.7 Model Discussion and the Gains from Openness

Offshore R&D and production by MNCs integrate countries into a four-stage GVC, starting fromheadquarter services embedded in firm knowhow, to R&D, production, and eventually to market-ing and consumption. When firms are allowed to carry out R&D abroad, participants at differentstages of the GVC benefit. First, firm owners and high-skill workers in the host benefit by sharingthe rent from the newly created varieties. Because manufacturing production tends to be closeto R&D, low-skill workers in the host and nearby countries benefit through increasing demandfor labor. Finally, high-skill workers in destination countries benefit from the increase in labordemand for marketing and everyone benefits from expanding access to varieties.

Given the interaction among these forces, characterizing the welfare implications of a generalpolicy on offshore R&D is challenging. To illustrate the intuition, I derive an expression for thewelfare effect of a particular counterfactual: moving the economy from the observed level of in-ternational linkages to complete isolation. For only this section, I make the following assumption:

Assumption 3. a) The distribution from which firms draw production efficiencies zP is independent of zR

but can vary arbitrarily across pairs of origin country and R&D location, o and i.b) Firm innovation efficiency, zR, follows a Pareto distribution: GE

o (x) = 1− ( xZR

o)−κR , with ZR

o being thecountry specific lower bound and κR > (1− γ) being the tail coefficient.c) Lh

i and Lli are exogenous, i.e., there is no occupation choice

Part a of the assumption departs from Assumption 2 as it rules out correlation between zP andzR at the firm level. On the other hand, zP can be from any parametric distributions and differacross pairs of origin and host countries, so heterogeneity at country-pair level or the effect ofdistance to headquarters on zP are incorporated. Parts b and c are technically assumptions thatsimply expressions. Under Assumption 3, we have the following:

Proposition 1. If Assumption 3 is maintained in place of Assumption 2, then the gains from opennessfor country d, defined as the percentage change in Xd

Pdas a country moves from complete isolation to the

observed equilibrium, is

GOd =(Xdddd

∑m Xddmd)−

1θ × (

∑o,m Xodmd

Xd)−

1θ × (

∑m Xddmd

∑o,m Xodmd)−

1θ × (

Idd

∑o Iod)

γθ × f (∑o Iod

Xd,

Idd

∑o Iod)× Xd

Yd− 1,

(9)

where f (∑o IodXd

, Idd∑o Iod

) is a function of model parameters and two ratios: the share of variable R&D expenses

in income ∑o IodXd

, and the fraction of domestic R&D expenses by local firms Idd∑o Iod

.

Proof. See appendix.

This expression illustrates forces through which a country benefits from economic integration.The first term, Xdddd

∑m Xddmd, measures the importance of trade. In the absence of colocation and head-

24

quarter effects, i.e., when φPoim is independent of both i and o, this term equal to ∑o,i Xdidd

∑o,i,m Xdimd, which

is the fraction of sales in country d that is manufactured locally, the same as in the familiar gainsfrom trade formula. The second term, ∑o,m Xodmd

Xd, is the fraction of country d spending on goods

invented in d. It captures the benefit from having access to goods invented else. Together, the firsttwo terms capture the direct effect of offshore production and trade.

The third term, ∑m Xddmd∑o,m Xodmd

, captures the importance of foreign firms in domestic R&D. Intuitively,the smaller this term is, the more country d depends on foreign affiliates for R&D and the moreit loses when these foreign R&D centers are shut down. This channel is intermediated by theability of domestic firms to make up for the lost varieties—the withdraw of foreign firms releasehigh-skill workers to work on R&D at local firms. Holding the total number of workers in productdevelopment constant, the strength of this offsetting channel is captured by the share of local firmsin R&D, ( Idd

∑o Iod)

γθ , the fourth term.

Together, the first four terms captures the direct impacts of openness on the real wage of low-skill workers. The last two terms account for the indirect influence of openness on the aggregatewelfare through changes in the sources of income. More specifically, Xd

Ydcaptures that openness

allows some countries to specialize in innovation and earn a higher share of income through in-novation rent. For such countries, Xd

Ydis higher and the welfare gains are above and beyond the

gains in the real wage. f (∑o IodXd

, Idd∑o Iod

) accounts for that the share in income from innovation andmarketing as well as the share of high-skill workers devoted to R&D can endogenously respondto openness.

In the appendix, I compare this formula to that from Arkolakis et al. (2018). The first two andlast terms in Equation (9) also appear in their formula, with adjustments to reflect the modelingdifferences. The third and fourth terms, ( Xddd

∑oXodd)−

1θ × ( Idd

∑o Iod)

γθ , which summarize the direct effect

of offshore R&D, is novel. For an idea of the importance of this effect, consider a special case withφP

oil independent of i, in which case ∑l Xddld∑o,l Xodld

= ∑i,l Xdild∑o,i,l Xoild

= Idd∑o Iod

, and the product of the third and

fourth terms simplifies to ( Idd∑o Iod

)−1−γ

θ . Consider the median country in the quantitative section,

with about 30% of its R&D done by foreign affiliates. The value of ( Idd∑o Iod

)−1−γ

θ is around 1.05,when γ = 0.25 and θ = 4.5—all else equal, offshore R&D brings about 5% direct welfare gains.

In addition to bringing about direct welfare gains, incorporating offshore R&D also affectsthe inferred gains from trade and offshore production. With R&D and production colocation, mymodel infers different values for the first two terms of the formula, which are usually not directlyobservable. In the next section, I will parameterize the full model and show that using the firm-level data is important for both quantifying the gains from offshore R&D and the other channelsof the gains from openness.

4 Parameterization

In quantifying the model, I focus on the same 37 countries as in the empirical analysis. I parame-terize the model to be consistent with the data in the following aspects: the estimated geographicfrictions from regressions in Section 2.3, country-specific measures of openness and income, and

25

firm-level statistics of firms from the U.S. This section describes the parameterization procedures,starting with the functional form assumptions.

4.1 Additional Assumptions

Talent and innovation management efficiency distributions. I parameterize the ability distribu-tion for workers in country i, Ai(α), to be a log normal distribution governed by µi

α and σiα (the

mean and standard deviation of the associated normal distribution, respectively). I parameterizefirm innovation efficiency to be from truncated Pareto distributions as below:

GEo (z

R) =(ZR

o )−κR

o − (zR)−κRo

ZRo−κR

o − ZRo−κR

o, (10)

in which the ZRo and ZR

o are the lower and upper bounds of the distribution and κRo is the tail

parameter of the distribution, all of which can vary across countries. The distribution convergesto the Pareto distribution as ZR

o → ∞.Relationship between zP and zR. To capture the correlation between innovation and produc-

tion efficiency, I assume that there are two distributions, denoted by GH(zP) and GL(zP), fromwhich firms draw their productivity zP. The probability of drawing from the high distributionGH(zP) depends on the innovation efficiency of an R&D center as follows:

Prob(zP ∈ H|zR) =exp(δ0 + δ1 × zR)

1 + exp(δ0 + δ1 × zR), (11)

Parameters δ0 and δ1 are to be estimated. A positive value for δ1 means that more innovative firmstend to be more productive as well. GP

H(zP) and GP

L (zP) are both Pareto distribution with the same

shape parameter κP:

GPH(z

P) = 1− (zP

HzP )κP , GP

L (zP) = 1− (

zPL

zP )κP .

I assume that zPL < zP

H, so GPH(z

P) first order stochastically dominates GPL (z

P).Geographic frictions. I parameterize frictions impeding offshore production and R&D as log

linear functions of various distance measures and a host-specific fixed effects as follows:

log(φP

oim) = s · log(φPim) + (1− s) · log(φP

om), s ∈ [0, 1],

log(φPom) = 1(o 6= m) · [φP

m +−−→βP,om ·

−−−→distom]

log(φPim) = 1(i 6= m) · [φP

m +−−→βP,im ·

−−−→distim]

(12)

log(φR

oi) = 1(o 6= i) · [φRi +−→βR ·−−→distoi]

cRoi = 1(o 6= i) · exp

(φcR

i +−→βcR ·

−−→distoi

)The first block of Equation (12) defines φP

oim, the efficiency retained in offshore production in

26

country m, as the geometric average of φPim and φP

om, with s capturing relative importance of thetwo terms. When s = 1, only proximity to R&D centers (φP

im) matter; when s = 0, only prox-imity to headquarters (φP

om) matter. φPom and φP

im are themselves functions of φPm, which captures

host-specific effects on inward offshore production, and bilateral distance measures. As in thereduced-form analysis, four distance measures are included: geographic distance and indicatorsfor whether the two countries share an official language, are contiguous, or have a colonial tie. Thesecond block in Equation (12) defines the retained efficiency and the fixed cost in doing offshoreR&D, φR

oi and cRoi. Through the inclusion of φR

i and φcRi , the parameterization allows openness to

foreign R&D to differ across hosts. Finally, the indicator functions in these definitions normalizeφP

om, φRoi, φP

im to 1 and cRoi to zero for own countries.

4.2 Parameters Assigned Directly

I directly assign values to some parameters of the model and this subsection explains how. Addi-tional details on the data used are provided in Appendix C.1.

Labor endowment. Because patenting and trade are by and large a manufacturing activity, Iinterpret the model as for the manufacturing industry. I set the number of workers in a country,Li, to the manufacturing employment of a country, calculated using total employment from thePenn World Table and manufacturing share from the World Bank.

Elasticities. The elasticity of substitution between varieties, σ, determines markups. I set thisparameter to be 4. This value implies a markups of 30%, in line with recent findings using firm-level data from the U.S. and elsewhere (De Loecker and Eeckhout, 2018). θ governs the dispersionof the idiosyncratic component of affiliate productivity. As Equation (3) suggests, the share ofsales devoted to marketing is θ−(σ−1)

θσ . Base on a recent survey of chief marketing officers,22 U.S.firms spend on around 8-10% of revenues on marketing. I set θ = 4.5, so the marketing budget is8.3% of sales in the model. This choice of θ is also close to the estimate off trade flows by Arkolakiset al. (2018). With θ and σ given, γ determines the share of sales spent on developing new varieties( γ(σ−1)

θσ ), which I interpret as the revenue share of R&D expenses in the data. U.S. manufacturingfirms in the Compustat data on average spend 4% of sales on R&D. Setting γ(σ−1)

θσ = 0.04 givesγ = 0.25. After deducting production cost, marketing, product development, the remaining com-ponent of sales is the variable profit. Accounting for about 12% of sales, this component coversboth the overhead in offshore R&D and the net return to firm knowhow.

Trade costs. I normalize τmm = 1 and assume that trade costs are symmetric. I show inAppendix C.2 that the approach of Head and Ries (2001) applies to this model, i.e., we can backout trade costs as:

τmd = τdm = (∑o,i Xoimd

∑o,i Xoimm· ∑o,i Xoidm

∑o,i Xoidd)−

12θ .

Although Xoimd is not observable, the four summations in the expression are just entires in con-ventional input-output tables. I aggregate across manufacturing industries in the World Input

22See https://cmosurvey.org/about/

27

Output Database and pin down bilateral trade costs using this equation.Firm efficiency and worker ability distribution. Calibrating firm efficiency and worker talent

distributions requires comparable data across countries. I use the World Management Survey byBloom et al. (2012) for calibrating the former and the cognitive test score data by Hanushek andWoessmann (2012) for calibrating the latter.

The World Management Survey provides firm-level management scores for a number of coun-tries. In the survey, interviewers rate each firm based on its talent management policy and produc-tion efficiency along various dimensions. The overall management score for a firm is then a sum-mary of these sub scores and has been shown to be strongly correlated with objective measures offirm performance (Bloom et al., 2012). The talent management score intends to capture whetherfirms follow good managerial practice for retaining and incentivizing its talent, so it closely mapsto the capacity of a firm in R&D, where talent plays a disproportionate role. I use it to calibrateinnovation management distributions. I obtain three distribution statistics of zR for each country:mean, standard deviation, and skewness, and pick the parameters governing the distribution inEquation (10) so that the model statistics match their empirical counterparts.23

I define production management score to be the average of targeting, operation, and monitor-ing scores, which captures the efficiency of a firm in carrying out manufacturing efficiency. Withthis, I estimate δ0 and δ1 in Equation (11). Specifically, I classify a firm as being from a high produc-tivity distribution if its production management score falls into the top 5% of all firms in the sam-ple (corresponding to about top 12% in the U.S.). I then estimate the relationship between a firm’sinnovation management score and the probability that it is from a high productivity distributionusing the Logit specification implied by Equation (11). This procedure, reported in Appendix C.2,determines δ0 = −5, δ1 = 0.21.24

For talent distributions, from the test score database I obtain the average cognitive score andshares of students reaching ‘basic’ and ‘top’ performance level, which are defined based on a com-mon standard across countries. I normalize the distribution of the U.S., assuming its talent distri-bution is standard log normal (i.e., µα

US = 1, σαUS = 1), and then determine talent distributions of

other countries by matching their distribution statistics relative to those of the U.S.25

Table 6 summarizes information on the parameters determined directly. I choose additional

23The talent management score in the data follows a symmetric bell-shaped distribution. Since it is well knownthat the firm size distribution has a fat tail, I calculate the three statistics based on the exponent of the original scores,effectively treating the reported scores as logarithm of actual innovation efficiency. Some countries in quantification arenot covered by the World Management Survey. I impute their statistics based on income and the geographic regions ofcountries. Appendix C.1 reports the distribution statistics for both firm knowhow and worker ability.

24The choice of the top 5% cutoff is motivated by the importance of the most productive firms in international busi-ness. A high cutoff allows me to better capture the activities of these firms. An implicit assumption in using the Logitregression to estimate δ0 and δ1 is that firms drawing their production efficiency from GP

L (zP) constitute the bottom

95% in the production efficiency distribution, whereas firms drawing from GPH(zP) constitute the top 5%. This assump-

tion does not hold exactly because under the Pareto assumption, GL(zP) will always overlap with GH(zP). Given thechoice of the cutoff (5%), however, the calibrated ZP

H will be large enough so the overlap is negligible. Figure C.1 in theappendix plots the probability of a firm drawing from GH(zP) against zR for the U.S.

25Specifically, in the U.S., 91% and 7.3% students reach ‘basic’ and ‘top’ performance, respectively, which impliesthat the cutoffs for basic performance is 0.25 and for top performance is 4.27; the U.S. distribution has a mean of 4.9. Icalibrate µα

i and σαi for other countries so that their shares of population above the two cutoffs and their mean scores

relative to that of the U.S. are consistent with the data.

28

Table 6: Parameters Calibrated Externally

Symbol Description Value Source

σ elasticity of substitution between varieties 4 markup (30%)θ dispersion of offshore production draws 4.5 marketing expenditures/sales (8.3%)γ researcher share of variable profit 0.25 RD expense/sales (4%)δ0 probability of high production efficiency -5 estimated (Table C.3)δ1 dependence of zP on zR 0.21 estimated (Table C.3)τmd|m, d = 1, ..., N trade costs - World Input Output DatabaseGE

o (zR)|o = 1, ..., N innovation efficiency dist. - Bloom et al. (2012)Ai(α)|i = 1, ..., N talent dist. - Hanushek and Woessmann (2012)Li|i = 1, .., N labor endowment - World Bank & PWT

parameters jointly in equilibrium, following the process described below.

4.3 Parameters Determined in Equilibrium

Overview. The remaining parameters to be determined include production efficiency distributionparameters, zP

L , zPH, and κP, country-specific productivity, Tm|m = 1, ..., N, measures of firms in

a country Eo|o = 1, .., N, frictions to offshore production and R&D, which include host-specific

components φPm, φR

i , φcRi |i, m = 1, .., N, the bilateral coefficients

−−→βP,om,

−−→βP,im,

−→βR,−→βcR, and the

weight parameter s. Although in equilibrium these parameters are jointly identified, for certainparameters some moments are more informative than others. I describe the intuition on how eachparameter is determined and the numerical algorithm used.

Firm production efficiency parameters. Having pined down distributions of innovation effi-ciency distribution and the way it maps into productivity, governed by Equation (11), the param-eters of GP

H(zP) and GP

L (zP) are the remaining degrees of freedom for the firm size distribution. I

normalize zPL to 1, and pick zP

H and κP jointly. κP directly affects the shape of the firm size distri-bution at the very top, while zP

H has a direct influence on the scale of the top 5% firms relative tothe rest. The targets are the Pareto tail coefficient for the firm size distribution, and the fraction offirms withe fewer than 20 and 100 employment, all of which are based on the U.S. data and willbe compared to the corresponding statistics of the U.S. in the model economy.

Measure of firms. The measure of firms from country o, Eo, along with their innovation effi-ciency, determines the fraction of the world patents invented by firms from o, ∑d Vod

∑o,d Vod, which has

a well-defined empirical counterpart. I normalize EUS and chose Eo : o 6= US so the model isaligned with the data.

Manufacturing efficiency. Tm determines the overall manufacturing efficiency in country m. Icalibrate Tm so that Xm

Pmmatches the real GDP of country m.

Geographic frictions to offshore R&D and production. I determine the three sets of host-specific parameters, φP

m, φRi , φcR

i |i, m = 1, .., N, by matching the overall openness of a countryto inward offshore R&D (both extensive and intensive margins) and inward offshore production.The moments that I target are the foreign share of patents invented in a host ∑o 6=i Voi

∑o Voi, the foreign

share of R&D center counts ∑o 6=i Roi

∑o Roi, and the foreign share of manufacturing production ∑o 6=m Yom

∑o Yom.

29

The empirical counterparts of these moments are aggregated from firm-level data used in Section2.2, with some adjustments described in the appendix. A comment on the target for the intensivemargin of offshore R&D. In the model ∑o 6=i Voi

∑o Voiis variety counts without accounting for the ‘quality’

of these varieties, captured by zP. Consequentially, the target I use is simple patent counts. Thatpatents from different countries could be of different values are implicitly incorporated throughthe distribution of zP.

Given φPm, φR

i , φcRi |i, m = 1, .., N, I adopt an indirect inference procedure to infer the remain-

ing parameters for geographic frictions, i.e., s and −−→βP,om,

−−→βP,im,

−→βR,−→βcR. Specifically, I target

reduced-form estimates from five specifications in Section 2.3 that are most informative aboutthese coefficients. The first three specifications are Columns 1, 2, and 5 of Table 4, which estimatethe headquarter effects for the intensive margin of offshore production, and for the extensive and

intensive margins of offshore R&D. These three regressions help identify−−→βP,om,

−→βR, and

−→βcR. The

remaining parameters,−−→βP,im and s, jointly determines the relative importance in production of ac-

cess to R&D centers versus headquarters. I choose two regressions identifying precisely this asthe target. First, Column 3 of Table 3. This specification controls for the home-host fixed effectsand thus directly identifies the effects of being close to offshore R&D centers. Second, Column 6 ofTable 4. Controlling for distance to headquarters, the coefficient for R&D center indicator conveysinformation on the importance of access to R&D.26

I simulate 50,000 firms, with the number from a country proportional to its size, and solve fortheir optimal R&D and production allocations. I then use the simulated data to estimate the fiveregressions above, including the exact same set of controls as in the corresponding specifications. I

choose −−→βP,om,

−−→βP,im,

−→βR,−→βcR and s to minimize the distance between regression coefficients based

on the simulated data and those based on actual data. With 22 target coefficients and 17 parame-ters, this procedure is over-identified. In constructing the objective function, I weight regressioncoefficients using the inverse of their empirical standard errors reported in Section 2.3, acknowl-edging that not all coefficients are estimated with the same precision.

Numerical implementation. The parameterization in this subsection follows a nested fixedpoint algorithm. In the outermost layer, I choose zP

H and κP to match moments on firm size. In

the middle loop, I search over the space of −−→βP,om,

−−→βP,im,

−→βR,−→βcR and s to minimize difference

in regression coefficients. In the inner most loop, I solve for the competitive equilibrium, whilechoosing Tm|m = 1, ..., N, Eo|o = 1, .., N, andφP

m, φRi , φcR

i |i, m = 1, .., N to match their re-spective targets described above. Once the innermost loop is completed, I simulate the model toobtain regression coefficients to feed into the middle loop.

The nested procedure requires solving the model efficiently and accurately, which involvesrepeated summations over a high-dimensional objects, Xoimd, and numerical integrations overheterogeneous firms for aggregate objects. The online Appendix provides details on the imple-

26Note s and−−→βP,im can be separated because I have imposed that the weights of φP

im and φPom sum to one. I do not

target the regressions where the patent count is an explanatory variable because these regressions use much smallersamples and are more likely to subject to the attenuation bias.

30

Table 7: Parameters Calibrated in Equilibrium

Parameter and value Description Moment Model Data

A. Firm size dist. parameterszP

L = 1 (normalized) % of firms with emp.<100 0.99 0.99zP

H = 2.01 Firm zP draws % of firms with emp.<20 0.94 0.95κP = 6.15 Power law coefficient

of firm size dist.1.05 1.05

B. Country-specific parameters and fixed effectsTm|m = 1, ..., N country-specific man. TFP Xm

Pm -

Columns5-10 ofTableC.1

Eo|o = 1, .., N measure of domestic firms ∑i Voi∑o,i Voi

-

φPm|m = 1, .., N host offshore production ∑o 6=m Yom

∑o Yom -

φRi |i = 1, .., N host offshore R&D efficiency ∑o 6=i Voi

∑o Voi -

φcRi |i = 1, .., N host offshore R&D overhead ∑o 6=i Roi

∑o Roi -

C. Bilateral Geographic Coefficientss=0.82 φP

oim = (φPim)

s(φPom)

1−s

distance lang border colonyreduced-form estimateson colocation andheadquarter effects

TableC.4Panel B

TableC.4Panel A

-0.063 0.014 0.045 0.022−−→βP,im

-0.032 0.038 0.022 0.054−−→βP,om

-0.061 0.17 0.043 0.026−→βR

0.12 -0.027 -0.10 -0.0052−→βcR

mentation of the algorithm.27

Parameter values and model fit. Table 7 summarizes the results from the joint calibration.Panel A is the parameters determined in the outermost loop. The calibration process determinesthat zP

H = 2 and κP = 6.15. Pining down these two parameters are three moments of the U.S. firmsize distribution, which the model matches closely.

Panel B is the country specific parameters that are just-identified and matched perfectly bydesign in the innermost layer. The target values of the moments are reported Appendix Table C.1.

Panel C of Table 7 reports the coefficients governing bilateral costs of offshroe R&D and pro-duction. With an s of 0.82, it implies that proximity to the R&D team is more important than prox-imity to the headquarters. This reflects strong co-location pattern documented in the empiricalsection. All distance coefficients have expected signs—geographic distance impedes the separa-tion of R&D and production from each other and from the headquarters, whereas sharing a border,a common language, or colonial ties encourages it. In line with the heterogeneous headquarter ef-fects by activity documented before, the distance coefficients are different across activities. Forexample, sharing an official language with the headquarters increases official R&D efficiency by17% but increases offshore production efficiency by only 3.8%. Appendix Table C.4 reports themodel and data moments side by side for comparison and shows that the model fits data well: allbut three of the 22 coefficients are within 95% confidence intervals of empirical estimates.

27An often used alternative is to cast the entire procedure as a constraint optimization problem, with all equilibriumconditions and the set of just-identified moments being the constraint. This alternative is not suitable here becauseeach evaluation of the constraint optimization problem requires estimating firm-level regressions, which is costly giventhat a large number of firms are simulated and many fixed effects are included. The nested approach circumvents thisproblem because it only perform this regression for when the inner loop are satisfied.

31

Table 8: Fit of Untargeted Moments

Additional moments on firm size in U.S. Model DataFraction of firms with emp. < 10 0.91 0.90Share of emp. in firms with emp. > 500 0.611 0.47Share of R&D by parents of MNCs 0.84 0.79

The efficiency advantage of foreign affiliatesForeign affiliate advantage 0.21 0.15coefficient of variation across countries 1.272 1.158correlation with log host GDP per capita -0.11 -0.25

4.4 Validation using Non-targeted Moments

I evaluate fit of the model on non-targeted moments. Table 8 reports the results.Employment and R&D concentration. The upper panel reports additional statistics on firm

size. In the data, about 90% of firms have fewer than 10 employees and about 47% of employmentare in firms with more than 500 employees. The model matches the former well and slightly overpredicts the latter. Turning into the concentration of R&D, in 2014, about 79% of the businessenterprise R&D is conducted by parent firms of U.S. multinationals. The model prediction falls inthe same ball park.

The multinational managerial advantage. The model assumes that knowhow is the source offirm heterogeneity driving offshore R&D decisions and disciplines the extent of heterogeneity us-ing management scores. This assumption, together with self-selection into offshore R&D, impliesthat foreign affiliates tend to have higher management scores than indigenous firms and that thisdifference is larger in low-income host countries populated with inefficiently managed firms.

To validate these implications quantitatively, I calculate the foreign affiliate managerial advan-tage for each country, defined as the percentage difference between the average innovation effi-ciency of foreign affiliates and that of indigenous firms, and compare the statistics of this measureto their data counterparts constructed using the World Management Survey.

As the lower panel of Table 8 shows, the model prediction of average foreign affiliate man-agement advantage and its cross-country variations are similar to those in the data. Moreover,both the model and the data show a negative correlation between this measure and host income,although the model under-predicts the strength of the correlation.

Bilateral offshore production and R&D. The parameterization matches the overall inwardoffshore R&D and production of each host by design, but when it comes to bilateral offshore pro-duction and R&D, the only information used in calibration are the regression coefficients esti-mated with firm and country fixed effects and thus identified only from within-firm variations. Itis therefore useful to see whether the model matches the data closely on aggregate bilateral off-shore production and R&D shares. Figure 3a shows that log offshore production shares matchesits empirical counterparts closely. Figure 3b shows that same is true for log offshore R&D shares,except for some country pairs on the left end, which have very little offshore R&D in the data. Forexample, a log offshore R&D share of −14 corresponds to a data point accounting for less than

32

Figure 3: The fit of Bilateral MP and offshore RD

(a) MP

-18 -16 -14 -12 -10 -8 -6 -4 -2 0

Data

-18

-16

-14

-12

-10

-8

-6

-4

-2

0

Model

corr = 0.73

(b) R&D

-16 -14 -12 -10 -8 -6 -4 -2 0

Data

-16

-14

-12

-10

-8

-6

-4

-2

0

Model

corr = 0.84

Notes: The left panel shows that the fit of the model in log bilateral multinational production shares, defined as log( Yom∑o Yom

), ∀o, m; the

right panel shows the fit of the model in log bilateral offshore R&D shares, defined as log( Voi∑o Voi

), ∀o, i.

1e-8% of host R&D.Occupation choice of workers. My parameterization matches the R&D by origin country of

firms and the real income by country, and therefore indirectly discipline shares of income fromhigh-skill workers and other sources. Whether the resulting sources of income translate into a rea-sonable occupation choice of workers boils down to if the log normal distribution, parameterizedto match cognitive test scores, is a good approximation for the reality. As a validation of these as-sumptions, Figure 4 plots the share of workers sorting into high-skill occupations, which includethe fixed and variable part of R&D and marketing, against the share of researchers in industrialemployment from the OECD. The range of variation in the data is larger than in the model, likelybecause the definition of skill is narrower in the data than in the model. Despite this difference,the correlation between the two variables is 0.75, showing that the model captures the variationsof the skill composition across countries well.

4.5 The Role of Offshore R&D for Production and Income

The parameterized model enables descriptive characterizations of various slices of the GVC thatare often difficult to measure directly. I focus on two aspects—the use of R&D by host country andthe compositions of firms’ profit—to demonstrate the important role of offshore R&D in shapingcountries’ integration in the GVC and their sources of income.

Columns 1 through 4 of Table 9 decompose R&D by its source (whether done by local firms)and use (whether used for local production). The first and third columns are the shares of R&Dby domestic and foreign firms, calibrated to perfectly match the data. On average, about a third ofR&D in these countries are conducted by foreign affiliates. This ratio tends to be higher in smallopen European countries and lower in global R&D powerhouses such as the U.S. and Japan.

The second and fourth columns are inferred through the lens of the model. The second column

33

Figure 4: Model Validation: High Skill Shares

-8 -7.5 -7 -6.5 -6 -5.5 -5 -4.5 -4

Data: log (research personnel/ Industrial employment)

-3.1

-3

-2.9

-2.8

-2.7

-2.6

-2.5

-2.4

-2.3

Model: log (

hig

h s

kill

share

)

AUSAUT

BEL

CANCHE

CZE

DEU

DNKESP

EST

FIN

FRA

GBR

GRC

HUN IRL

ITA

JPN

KOR

LTULVA

MEX

NLD

NOR

POL PRT

ROU

RUS

SVK

SVN

SWE

TUR

USA

corr = 0.75

Notes: The vertical axis is log of high-skill occupation share in the model, i.e., log(1− Ai(ai)). The horizontal axis is log of the shareof researchers in industrial employment obtained from the OECD.

is the local production share of varieties developed by domestic firms at home. I measure thisshare by revenue of the varieties developed, i.e., ∑d Xoood

∑m,d Xoomd. This share averages 82.9%, reflecting

that it is costly to move production away from the headquarters and the R&D center at the sametime. The fourth column reports the local production share of varieties developed by foreignaffiliates. On average, 70% of inward offshore R&D leads to offshore production in the same host,so access to host producers is important. Figure 5a plot the local production share of foreignaffiliate R&D against the average cost for manufacturing production, approximated by log(W l

mTm

).Although the fit is far from perfect, most of the variations in the data can be explained by myproduction cost proxy—offshore R&D in low-income countries are more likely to lead to localproduction than that in high-income countries. The large variations in local production sharesand its systematic correlation with country characteristics further demonstrate that we cannotmechanically bundle R&D and production in a single location chosen by each firm.

Offshore R&D enables firms to apply their knowhow globally and shape the sources of in-come for all. Columns 5 through 9 break down the total income of a country into manufacturingproduction, profit (total and that made from overseas R&D), R&D, and marketing. In autarky,the sources of income are the same across countries. In the open economy, while the marketingexpenditure share is still a fixed share of income, the importance of the remaining components arealtered by firms’ global operations. Advanced countries populated with the most efficient firmstend to earn a higher share from profit. For example, profit accounts for 21% of income in the U.S.,among which more than a third is from overseas R&D centers. In contrast, only 5% of income inSlovakia is from profit and almost all of it is generated from varieties invented domestically. Therelative abundance of a country on firm knowhow versus talented workers governs firms’ incen-tive to combine knowhow and talent from different countries, and therefore has a direct impacton income compositions, especially the ratio between profit and R&D income. In models withoutoffshore R&D, this ratio is γ

1−γ = 3, in my model this ratio ranges 1.7 to 6. Figure 5b plots theincome ratio from profit over R&D. Countries that are relatively more abundant in firm knowhow

34

Table 9: Decomposition of R&D and Sources of Firm Profit

(1) (2) (3) (4) (5) (6) (7) (8) (9)Source and use of R&D Source of income (% of total income)

% by domestic firms % by foreign firms manu. profit R&D mkt.

% of localprod.

% of localprod.

totalinventedabroadCountry

AUS 60.2 87.3 39.8 73.5 82.6 6.4 0.24 2.6 8.3AUT 54.5 78.8 45.5 60.1 82.1 6.8 0.44 2.7 8.3BEL 41.1 63.6 58.9 45.9 85.0 4.6 0.72 2.0 8.3BGR 81.0 97.2 19.0 87.4 79.2 9.2 0.01 3.2 8.3BRA 42.4 97.4 57.6 84.1 76.6 10.6 0.01 4.5 8.3CAN 47.8 84.9 52.2 67.1 75.5 11.9 2.37 4.3 8.3CHE 58.4 71.3 41.6 50.3 74.1 13.2 2.16 4.4 8.3CHN 58.5 95.0 41.5 74.3 76.7 10.2 0.00 4.8 8.3CZE 72.7 92.7 27.3 80.7 83.8 5.7 0.01 2.2 8.3DEU 71.8 70.5 28.2 59.4 76.9 11.2 1.69 3.6 8.3DNK 61.6 66.5 38.4 44.9 79.3 9.6 2.40 2.8 8.3ESP 77.9 90.8 22.1 79.5 81.4 7.6 0.09 2.7 8.3EST 71.5 93.2 28.5 75.4 82.3 6.9 0.05 2.4 8.3FIN 75.1 73.0 24.9 47.6 73.0 14.2 2.23 4.5 8.3FRA 73.4 71.0 26.6 55.8 76.4 11.6 1.39 3.7 8.3GBR 37.0 46.8 63.0 39.0 83.9 5.4 1.99 2.4 8.3GRC 49.9 97.3 50.1 87.5 77.9 10.1 0.02 3.7 8.3HRV 46.6 98.3 53.4 90.4 78.5 9.5 0.00 3.6 8.3HUN 44.2 97.7 55.8 89.2 80.4 7.2 0.00 4.0 8.3IRL 26.4 47.3 73.6 32.5 70.5 17.5 12.24 3.7 8.3ITA 56.1 92.9 43.9 80.9 78.2 9.4 0.14 4.1 8.3JPN 97.7 83.9 2.3 68.0 72.1 15.1 1.75 4.5 8.3KOR 93.4 90.0 6.6 75.6 74.4 13.1 0.65 4.2 8.3LTU 79.7 94.8 20.3 81.3 81.3 7.7 0.11 2.7 8.3LVA 94.0 97.2 6.0 86.0 80.3 8.5 0.01 2.9 8.3MEX 45.0 98.9 55.0 94.7 76.9 10.7 0.01 4.1 8.3NLD 69.7 14.9 30.3 12.3 80.5 9.0 3.60 2.2 8.3NOR 68.3 68.9 31.7 45.1 67.7 20.4 11.80 3.5 8.3POL 85.4 92.1 14.6 80.9 81.6 7.5 0.11 2.5 8.3PRT 73.8 94.9 26.2 83.0 81.3 7.7 0.01 2.7 8.3ROU 46.5 98.5 53.5 94.4 84.3 5.1 0.00 2.3 8.3RUS 88.4 91.7 11.6 81.6 79.1 9.4 0.06 3.2 8.3SVK 71.6 94.5 28.4 82.8 85.1 4.8 0.02 1.7 8.3SVN 82.7 94.7 17.3 79.5 79.7 8.9 0.02 3.1 8.3SWE 56.7 74.9 43.3 53.8 77.9 9.7 0.78 4.0 8.3TUR 79.9 97.9 20.1 89.3 75.4 12.0 0.00 4.2 8.3USA 84.2 67.6 15.8 58.3 66.0 20.8 7.65 4.9 8.3

Mean 65.6 82.9 34.4 69.5 78.3 10.0 1.5 3.4 8.3Notes: All numbers in this table are in percentage points. Columns 1 and 3 report the source of R&D, i.e., whether it isby domestic firms (Column 1) or foreign firms (Column 4). Column 2 reports the use of R&D by whether it is for localproduction. Columns 5 to 9 decompose the sources of income from different activities for a country: manufacturingproduction, profit (total and that accrued from products developed offshore), R&D, and marketing.

35

Figure 5: The Use of Offshore R&D and Its Influence on Income Compositions

(a) Local Production Share for Inward R&D

0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 20.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1

Inw

ard

offshore

R&

D local pro

duction s

hare

AUS

AUT

BEL

BGR BRA

CAN

CHE

CHN

CZE

DEU

DNK

ESP

EST

FIN

FRA

GBR

GRCHRV

HUN

IRL

ITA

JPNKOR

LTU

LVA

MEX

NLD

NOR

POL

PRT

ROU

RUSSVK

SVN

SWE

TUR

USA

(b) Ratio Between Profit and R&D Income

Notes: The left panel plots the share of inward offshore R&D devoted to local production against log average production cost. Theright panel plots the ratio between profit and researcher income against the relative abundance of firm knowhow, measured as log ofthe ratio between stock of firm knowhow and the stock of worker ability. The stock variables are constructed as the product of themeasures of firms (workers) and their mean knowhow (ability).

than in talented workers derive a higher share of income from profit than from R&D.The characterization of the GVC from two aspects in this section shows that offshore R&D

is an integrate part of firms global production decisions and has a first-order impact on the in-come distribution for all, but is purely descriptive. In the next section, I conduct counterfactualexperiments to isolate motives of offshore R&D and to throw light on its normative implications.

5 Counterfactual Experiments

5.1 Assessing the Two Motives of Offshore R&D

My first exercises focus on the forces driving offshore R&D. The model introduces two motives—arising from host relative talent abundance and its access to markets—for firms pondering whetherto enter a host. To assess the quantitative relevance of these factors, I conduct a set of experimentsin which I either change the distributions of talent and firm knowhow, or the market access of acountry. To isolate the effects from changes in other countries, when computing these counterfac-tual equilibria, I change parameters for one single country at a time, keeping model parameters atthe benchmark for all other countries.

The role of endowment distributions. To quantify the role of relative factor endowment, foreach host, I increase the mean of its innovation efficiency distribution and decrease mean of itstalent distribution, first separately and then jointly. Both changes reduce the incentive to enter thehost by weakening its relative talent abundance. For the changes to be realistic, I set the increase inthe mean innovation efficiency (achieved through increasing ZR

o ) and the decrease in mean talent(achieved through decreasing µα

i ) to be equal to the interquartile range of the mean knowhow andmean talent of the sample countries, respectively.

36

Table 10: The Determinants of Offshore R&D

“Talent Acquisition” “Market Access” All

Country Benchmark Knowhow Talent Both Consumer Producer Both(1) (2) (3) (4) (5) (6) (7) (8)

AUS 39.77 26.42 29.21 13.38 56.10 25.43 25.37 0.42AUT 45.51 11.95 30.93 2.33 68.22 2.50 2.46 0.06BEL 58.91 27.05 49.07 10.43 78.93 1.88 1.88 0.00BGR 19.01 1.29 4.66 0.00 30.40 8.60 8.60 0.00BRA 57.57 41.19 33.92 1.88 68.02 47.72 47.72 0.74CAN 52.21 40.43 43.22 27.61 57.17 34.08 33.48 0.00CHE 41.60 30.81 31.84 19.12 47.11 8.46 7.74 0.34CHN 41.48 36.40 31.77 26.08 50.15 34.41 34.41 16.49CZE 27.26 6.09 7.02 3.20 70.16 7.03 7.03 0.33DEU 28.20 22.54 22.02 16.88 32.40 18.51 18.32 6.00DNK 38.45 4.10 20.01 2.57 63.29 0.56 0.46 0.00ESP 22.10 1.30 4.28 0.51 42.92 3.57 3.60 0.00EST 28.51 15.29 16.45 5.39 85.51 14.43 14.43 0.00FIN 24.92 2.68 5.78 1.42 36.99 0.65 0.65 0.00FRA 26.60 21.96 19.71 15.55 30.38 16.79 16.62 6.29GBR 62.96 56.37 56.63 49.62 72.13 32.72 31.56 2.69GRC 50.08 4.82 4.04 2.10 64.55 31.43 31.43 0.56HRV 53.37 24.71 20.17 7.43 76.34 32.06 32.06 1.21HUN 55.81 39.21 37.64 26.30 79.43 44.66 44.66 20.29IRL 73.60 67.29 66.73 59.54 76.24 20.43 15.92 0.37ITA 43.88 22.41 24.73 2.72 59.84 29.33 29.26 1.58JPN 2.29 1.70 1.59 1.07 2.38 1.39 1.39 0.37KOR 6.57 2.69 2.96 0.10 8.42 2.56 2.52 0.02LTU 20.30 3.67 7.47 0.00 47.57 6.51 6.51 0.00LVA 6.04 0.93 1.59 0.00 8.41 2.58 2.58 0.00MEX 54.99 1.43 4.90 0.00 66.42 48.31 48.29 0.00NLD 30.27 19.76 24.31 14.49 41.38 0.00 0.00 0.00NOR 31.65 3.94 4.74 2.25 43.33 0.00 0.00 0.00POL 14.63 1.79 2.13 0.88 50.67 2.24 2.24 0.12PRT 26.22 2.38 2.00 0.62 52.65 2.79 2.79 0.00ROU 53.45 28.70 31.77 16.54 72.33 40.15 40.15 12.21RUS 11.59 0.82 1.07 0.40 26.06 0.96 0.96 0.13SVK 28.39 7.84 10.80 4.70 79.59 10.42 10.42 0.22SVN 17.29 6.71 5.79 1.07 69.91 6.86 6.86 0.00SWE 43.25 26.92 31.66 12.97 55.57 3.15 3.11 0.73TUR 20.11 1.96 1.27 0.23 33.97 3.93 3.93 0.00USA 15.77 12.92 12.61 10.10 15.91 10.82 10.80 5.91

Mean 34.4 17.0 19.1 9.7 51.9 15.1 14.9 2.1std 17.6 16.8 16.9 13.5 21.8 15.2 15.2 4.6

Notes: The numbers reported in this table are the share of domestic R&D expenditures incurred by affiliates of foreigncompanies in each country. All numbers are in percentage. The first column is from the baseline equilibrium. Columns2 to 4 changes firm knowhow distributions and worker ability distributions first separately and then jointly. Columns5 to 7 shut down export and offshore production from the host first separately and then jointly. Column 8 combines allchanges.

37

Column 2 in Table 10 reports the share of R&D by by foreign affiliates for a host after an im-provement in domestic firm innovation efficiency. As domestic firms become more competitive,wages increase and prices decrease, reducing profitability of foreign entry. Indeed, compared tothe baseline equilibrium, the foreign R&D shares in these counterfactual equilibria decrease byaround half to an average of 17%. The decreases are smaller for large countries like U.S., China,and Japan, perhaps because for them the product market consideration is more important. Col-umn 3 reports the results when the talent distribution of a country is worsened. On average,foreign affiliates account for about 20% of domestic R&D, marking a 40% decrease from the base-line. Intuitively, when domestic talent distribution deteriorates, R&D outputs of both domesticand foreign-affiliated R&D centers decrease. The decrease in the latter is more pounced becauseunable to recoup fixed R&D overheads, some choose not to enter at all. Finally, I simultaneouslychange the ability and innovation management efficiency distributions of a country. As Column 4of Table 10 shows, the average share of R&D by foreign firms drops to 10%, less than a third of thebenchmark value. This result shows that the calibrated international differences in distributions ofendowments can explain a significant share of the observed offshore R&D, especially for smallereconomies.

The role of host country market access. In the model, foreign access consists of two compo-nents: access to foreign consumers through exporting and access to foreign manufacturers throughoffshore production. I consider their separate and joint impacts.

In the first experiment, I increase the iceberg export cost of host m, τmd, m 6= d, to infinity. Thisshuts down host countries’ direct access to foreign consumers, but R&D centers there can stillaccess foreign consumers indirectly through offshore production. The shares of R&D by foreignaffiliates in these counterfactuals are reported in column 5 of Table 10, which show an across-the-board increase in offshore R&D shares.

This result might seem surprising at first glance, given the following partial equilibrium in-tuition: eliminating access to foreign consumers through direct exporting reduces the return todoing R&D in a host. This effect should reduce foreign firms’ incentive to enter as they have topay the overhead costs. In a model with both trade and offshore production, however, this directchannel is dampened—unable to export, firms can still serve foreign consumers by outsource pro-duction to other countries. Moreover, the reduction in export lowers production wages, pushingmarginal workers into R&D and making the country more attractive as a host for R&D centers.An increase in export costs thus has a similar effect to a decrease in a host country’s productionefficiency, which strengthens its comparative advantage in innovation and drives it to specializein R&D activities.

In the second experiment, I increase the cost of offshore production from an R&D host countryto infinity (by setting φP

oim = 0, i 6= m), so anything invented in i can only be produced locally.Column 6 of Table 10 shows that, most countries experience a decrease in offshore R&D fromthe baseline equilibrium. Unlike in the previous experiment, the general equilibrium effect actsin the same direction as the partial equilibrium effect: when the option of offshore production iseliminated, R&D centers in the host countries have to produce locally to serve both foreign and

38

domestic customers, which increases wages for production workers and researchers, making thecountry less attractive as a host for R&D centers. An increase in offshore production costs can beviewed similarly to a reduction in R&D innovation efficiency of a country, which strengthens itscomparative advantage in production.

I eliminate the two types of market access at the same time (Column 7). The inward offshoreR&D in this case is much smaller than when only export is shut down (Column 5), because whenoffshore production is not an option, countries can no longer specialize in innovation. It appearsthat once this specialization channel is eliminated, the partial equilibrium effect of export on off-shore R&D matters little—with the exception of Ireland, most numbers in Columns 7 are barelydifferent than in Column 6. When market access are eliminated and relative endowment differ-ences are reduced at the same time (Column 8), most countries attract very little foreign R&D.In summary, the two main forces incorporated in the model are able to account for most of theobserved offshore R&D.

5.2 Offshore R&D and the Welfare Gains from Openness

Having shown how offshore R&D responds to the two key forces in the model, I turn to its nor-mative implications. I examine the welfare gains from various forms of economic integrationsby eliminating each channel from the model separately. I define the gains from offshore R&D asthe increase in real income as whole model economy moves from an equilibrium where offshoreR&D is not allowed to the baseline equilibrium, and define the gains from trade and offshore pro-duction analogously. I compare these gains from individual channels to the overall gains fromopenness.

The gains from offshore R&D. The first column in Table 11 is the gains from offshore R&D.The simple average across countries is 3.3%, roughly the same as the gains from trade. This av-erage, however, mask important heterogeneity. While all countries benefit, advanced countries,such as the U.S. and western European countries, benefit the most. Emerging economies such asMexico, Russia, and China, benefit less.

What determines the heterogeneous gains? As discussed in Section 4.5, offshore R&D alters thesources of income most directly through redistributing rent from R&D across countries, leading toadditional profit for headquarter countries and additional researcher income in host countries. Imeasure the importance of this channel using the share of total innovation rent—the sum of profitand researcher income in a country—that is generated through offshore R&D, i.e., the income ofdomestic researchers working at a foreign affiliate plus the profit of domestic firms from overseasR&D. Figure 6a shows this measure indeed explains the variations in the gains from offshore R&D.

Interaction among the three forms of globalization. The second and third columns report thegains from trade and offshore production, respectively. The average is around 3.3% for trade, and8.4% for offshore production. As expected, smaller countries and those closer to major markets(e.g., Belgium and Netherlands) gain more from both channels, whereas large and more remotecountries (e.g., Brazil) gain less. In the fourth column are the gains from openness. Ranging from2.2% for Brazil to 82% for the Netherlands, it has an average value of 15.6%.

39

Table 11: Offshore R&D and the Gains from Openness

Baseline model Re-calibrated Alternative models

No Off. R&D s = 0 (or φPoil = φP

ol)

Country Off. R&D Trade Off. prod. Openness Openness Off. R&D Openness(1) (2) (3) (4) (5) (6) (7)

AUS 2.8 0.6 8.0 11.3 9.1 1.7 14.1AUT 2.8 4.0 10.1 17.4 15.3 4.2 24.0BEL 4.8 5.6 21.1 33.8 27.6 4.0 39.4BGR 0.6 2.8 2.3 6.1 5.9 13.1 22.3BRA 0.7 0.7 0.8 2.2 1.6 7.0 10.4CAN 4.1 3.7 5.5 12.0 8.1 11.0 26.1CHE 3.0 4.6 9.1 15.8 13.7 6.5 25.8CHN 0.9 0.5 1.1 2.7 1.9 1.9 5.3CZE 2.1 3.7 6.6 14.1 12.9 6.6 23.8DEU 3.0 4.4 10.1 17.0 15.2 4.1 23.2DNK 4.6 4.3 13.9 22.9 19.6 6.0 32.1ESP 1.4 1.4 5.5 8.7 8.2 5.2 15.9EST 1.5 3.8 5.4 11.7 11.5 8.6 24.3FIN 2.7 4.7 8.5 14.8 13.1 4.7 22.7FRA 2.1 3.5 9.6 14.8 14.1 3.6 21.0GBR 7.6 1.7 29.7 41.9 26.8 3.4 43.4GRC 0.5 1.4 1.6 3.7 3.4 14.4 20.5HRV 0.8 2.1 1.4 4.5 3.9 18.6 27.0HUN 2.9 6.0 2.2 11.3 8.2 13.6 31.5IRL 22.0 9.0 22.0 56.8 24.9 13.6 72.3ITA 1.4 2.2 2.7 6.2 5.0 8.7 17.5JPN 1.9 2.1 4.8 7.4 6.2 3.5 13.9KOR 0.8 2.5 2.9 5.6 5.3 4.8 13.0LTU 1.3 3.2 4.6 9.5 9.2 12.0 26.4LVA 0.7 2.9 3.1 7.2 7.3 14.1 25.6MEX 0.7 2.1 0.8 3.4 2.8 13.7 18.6NLD 8.0 6.1 60.0 82.6 75.2 1.8 74.6NOR 16.9 7.1 11.1 34.1 15.2 21.7 56.2POL 1.2 2.4 5.5 9.6 9.6 8.1 20.9PRT 1.1 1.4 4.3 7.3 7.0 7.5 17.1ROU 3.1 0.8 3.8 10.7 7.7 11.7 31.5RUS 0.7 1.2 3.6 5.8 5.9 5.1 12.7SVK 2.3 4.3 8.3 17.0 16.6 8.9 32.6SVN 0.7 3.8 3.1 8.1 8.1 9.9 20.9SWE 2.4 4.3 8.8 14.8 13.1 1.7 18.8TUR 0.1 1.8 0.7 2.5 2.4 14.1 17.7USA 9.1 5.2 9.6 22.4 12.7 4.8 26.6

mean 3.3 3.3 8.4 15.6 12.3 8.2 26.2std 4.4 1.9 10.6 16.0 12.3 4.9 14.6

Notes: All numbers are in percentage points. Columns 1 to 4 reports the gains from offshore R&D, trade, offshoreproduction, and openness, respectively. Column 5 reports the gains from openness in a restricted model withoutoffshore R&D, calibrated to match the same patterns of trade and offshore production as in the baseline equilibrium.Columns 6 and 7 report the gains from offshore R&D and openness for a re-calibrated that assumes that s = 0, in whichcase there is no colocaiton and all R&D is for production at the headquarters.

40

Figure 6: The Gains from Offshore R&D and Interaction with Trade and Offshore Production

(a) Gains from Offshore R&D (b) Interaction Among Different Forms of Integration

Notes: The left panel shows that the gains from offshore R&D (vertical axis) are higher in countries where offshore R&D account fora higher share of the innovation rent. The innovation rent is the sum of researcher income and profit in a country, and the fractionaccounted by offshore R&D is the profit of domestic firms doing R&D abroad plus the income of domestic researchers working inforeign R&D centers. The right panel shows that the three forms of international integrations are more complementarity to each otherin countries where purely domestic innovation rent (profit and researcher income generated without foreign engagement) accountsfor a higher share of income. The vertical axis is the ratio between the sum of individual gains and the gains from openness.

We can understand the interaction among the three forms of integration by comparing thesum of the gains from these individual channels to the gains from openness. If the former islarger, it means the benefit of further openness is greater once a country is already open in otherdimensions, so the three channels are complements for welfare; conversely, the three channels aresubstitutes.28 This measures shows that the three channels are complements for some (e.g., Japanand the U.S.) and substitutes for others (e.g., Romania).

This heterogeneity arise due to the specialization of countries in the world economy. By sep-arating where varieties are generated from where they are manufactured, trade and offshore pro-duction allows some countries to specialize in R&D and others in production. With efficient firmsmobilizing innovation knowhow globally, offshore R&D increases the effective efficiency of R&Din all countries. This enhances the comparative advantage of countries specializing in innova-tion but—as foreign R&D centers bid up domestic wages—weakens the comparative advantageof those specializing in production. As a result, the three channels are complements for the formergroup of countries and substitutes for the latter.

I use as a proxy for the comparative advantage of a country in innovation the share of itsincome from purely domestic innovation rent—the profit of varieties developed by domestic re-searchers with domestic firms only. This proxy captures the combined effect on the comparativeadvantage in innovation from both firm knowhow and domestic worker distribution. Figure 6bplots this proxy against a measure of complementarity—the ratio between the sum of gains fromindividual channels and the gains from openness—and shows that they are indeed strongly cor-

28Note that the gains from individual channels are calculated from the calibrated equilibrium, with the other twochannels both open. If the sum of individual channels are larger than the gains from openness, it means for at least oneof the channels, the effect of integration is larger when the other two channels are already open.

41

Figure 7: Geographic Frictions and the Gains from R&D and Openness

(a) Setting s = 0 redistributes the gains from off. R&D (b) Setting s = 0 amplifies the gains from openness

Notes: The left panel plots log of the ratio between the gains from offshore R&D in the alternative with s = 0 and the baseline model,against log of effective production cost. The right panel plots log of the ratio between the gains from openness in these two modelsagainst production cost.

related.Impact on the gains from openness. How important is offshore R&D for inferring the gains

from openness? Column 5 of Table 11 are the gains from openness calculated from a restrictedmodel without offshore R&D, calibrated to match the same trade and offshore production as inthe baseline equilibrium. This experiment generates 12.3% average gains from openness.

Compared to the restricted model, by incorporating offshore R&D, the baseline model ampli-fies the gains from openness by a factor of 1.3 (15.6/12.3) on average. This amplification differssignificantly across countries. Advanced countries earning a substantial profit from offshore R&Dsee the biggest increase in inferred gains. For example, incorporating offshore R&D almost dou-bles the inferred gains from openness of the U.S. For emerging countries on the receiving endof offshore R&D, the amplification is generally smaller and could even be slightly negative (e.g.,Russia). The uneven changes between baseline and restricted models further underscore the im-portance of incorporating offshore R&D—its omission not only underestimates the gains fromopenness, but also biases the comparison of the gains across countries.

The role of geographic frictions. In the model, the cost of offshore production to be a com-bination of distances to headquarters and distances to R&D centers. As discussed previously, myestimates suggest s = 0.82 and implies on average 70% of R&D in offshore centers is devoted toproduction in the same host. Now I show that the value of s, which I pin down using firm-leveldata, is crucial for correctly inferring the welfare gains.

To this end, I calculate the gains from R&D and openness in a re-calibrated model with s = 0,under which a higher fraction of offshore R&D is conducted for production at the headquarters.The last two columns of Table 11 report the results. The average gains from offshore R&D in-creases to 8.2% from the benchmark value of 3.3%, but the increase are concentrated in emerging

42

countries. In fact, the gains from offshore R&D for the U.S. shrink by almost half.The difference arises because the baseline model infers R&D as for mostly production in host

countries, which competes with local firms for manufacturing workers and product market, whereasthe alternative model infers it as mostly for headquarter production, which competes with otherfirms from the home country. Because developing countries with lower production costs have ahigher local production share of inward offshore R&D, moving to the alternative model with s = 0leads to a more significant reduction in competition faced by their domestic firms and, as a result,a larger increase in the gains from offshore R&D. The opposite is true for developed countries.Figure 7a plots the log change in gains from offshore R&D from the baseline to the alternativemodel with s = 0 and shows that indeed countries with higher effective production cost sees theinferred gains from offshore R&D decreased. Figure 7b further plots the changes in the overallgains from openness from the baseline to the alternative model. The negative correlation betweenproduction cost and changes in inferred gains continue to hold. Different from the gains fromoffshore R&D, however, the change are positive for all countries. This across-the-board increasein the inferred gains occurs because the alternative model implies a higher degree of integrationthrough offshore production. Recall that in calibration I match by country the inward offshore pro-duction share, ∑o 6=m Yom

∑o Yom. A substantial fraction of offshore production in this share is for varieties

developed locally by foreign R&D centers. Assuming offshore R&D as mostly for production atthe headquarters, the alternative model matches the same ∑o 6=m Yom

∑o Yomby allowing more production of

varieties developed elsewhere, effectively increasing the inferred openness and hence generatinghigher gains for all.

Taking stock, the counterfactual experiments in this subsection demonstrate that offshore R&Drepresents a new and quantitatively relevant channel through which countries benefit from glob-alization. It is a substitute for trade and offshore production for poor countries but a complementfor advanced countries. Further, the nature of offshore R&D—whether its inventions are devotedto production in the host or elsewhere matters a lot for the gains from offshore R&D and openness,stressing the need of using micro data to discipline the model.

5.3 Implications for the Global Incidence of FDI and R&D Policies

Most existing quantitative research on multinational activities does not differentiate policies onoffshore R&D and offshore production. Yet policy makers usually have at their disposal instru-ments that specifically target production or innovation activities. I examine whether not differ-entiating R&D and production is an important restriction for practical policy evaluations. As anexample, I focus on the further integration through FDI among emerging countries, which hasgained significance in the past decade.

Integration among emerging countries. I first consider an inward offshore R&D liberalizationthat eliminates the overhead cost for offshore R&D between a set of emerging countries, includingBrazil, China, Hungary, Mexico, Poland, Russia, Romania, and Turkey. This reduction in cost canbe interpreted as speedy approval of entry, or more broadly as subsidized land or tax credits for theupfront investment in R&D. The first column of Table 12 reports the results. Not surprisingly, this

43

policy benefits emerging countries. Yet their benefits are at the expense at developed countries,whose overseas R&D centers in emerging countries have to face tougher competition after thepolicy.

The second experiment is liberalization in bilateral offshore production, which increases φPim

by 20% between the set of emerging countries. I focus on the different distributions of the welfaregains across countries, rather than the level of welfare gains, because these two types of liber-alization do not necessarily have the same administrative burden or fiscal costs. As shown inColumn 2 of Table 12, emerging countries still benefit significantly, but differently from the firstexperiment, major developed countries are also better off. Allowing varieties developed in Russiato be more efficiently produced in China not only benefits Russia and China, but also Americanfirms doing R&D in Russia, who now make more profit. These two experiments demonstrate thatopenness to offshore R&D and offshore production could have different welfare implications forother countries.

Because trade and offshore production together form a GVC with offshore R&D, incorporatingthe latter also affect the effects of policies on trade and offshore production. To make this point,I consider the same liberalization as in the second experiment, but in a restricted version of themodel without offshore R&D. The welfare impacts of this experiment are reported in the thirdcolumn of Table 12. Compared to the baseline economy, emerging countries generally benefitless—without foreign R&D centers responding, the overall R&D in these host does not expandas much to take advantage of the increasing access to overseas producers. Developed countriesexperience net losses—the production of their affiliates in emerging countries see an increase incompetition as before, but now they cannot make up for the losses through developing R&D inthese countries. The comparison between these two experiments show that even if one’s goal issolely to understand the effect of liberalizing offshore production, it is important to incorporateoffshore R&D into the model.

The global impacts of R&D policies. The presence of offshore R&D also implies that R&Dpolicies can have a global impact simply because such policies would typically also apply to lo-cal R&D centers owned by foreign firms. This channel is independent of and in addition to thespillover effects of R&D across affiliates studied in the literature (e.g., Bilir and Morales, 2020),but can be just as important. As an example, I consider a 20% increase in R&D efficiency in GreatBritain. This change increases the real income of the Great Britain by 5.3%. Importantly, countrieswith extensive ties with Britain via offshore R&D also benefit. In total, 14% the total gains accrueto other countries. On the other hand, if offshore R&D is shut down, the same change in the R&Defficiency of the Great Britain will benefit itself by 5.7%, which is 133% of the total gains—othercountries, most notably the U.S., bear welfare losses.

6 Conclusion

Talented researchers and efficient firms are both necessary inputs to invention of products, butthey are distributed unevenly across countries. In an effort to take advantage of this uneven distri-

44

Table 12: Implications for FDI and R&D Policies

FDI Integration Among Emerging Economies Higher U.K. R&D Efficiency

off. R&D off. Prod.off. Prod. w/o.off. R&D

Baseline w/o off. R&DCountry

BRA 0.9 1.1 0.8 0.0 0.0CHN 1.1 1.2 1.3 0.0 0.0HUN 8.1 9.1 4.2 0.0 0.0MEX 0.7 1.9 0.8 0.0 0.0POL 1.6 7.9 3.9 0.1 0.0ROU 10.0 19.8 10.3 0.1 0.1RUS 2.5 4.7 2.7 0.0 0.0TUR 1.0 1.5 0.9 0.0 0.0

DEU -0.1 0.3 -0.1 0.0 0.0FRA -0.1 0.3 -0.1 0.0 0.0GBR 0.0 0.3 -0.1 5.3 5.7NLD 0.0 0.8 -0.2 0.4 0.2JPN -0.1 0.5 -0.2 -0.1 0.0USA -0.6 1.4 -0.3 0.1 -0.1

mean (all) 0.5 1.4 0.6 0.2 0.2Notes: The first two columns show that offshore R&D and production policies between the same set of countries have qualitativelydifferent incidences on the rest of countries. Comparison between Columns 2 and 3 show that the same offshore production policyhave different effects when offshore R&D is not incorporated. The last two columns shows that the spillover effects on the rest of theworld of an increase in R&D efficiency in U.K. are different if offshore R&D is overlooked.

bution in a world separated by geographic frictions, MNCs forge a GVC that integrate participantsfrom different parts of the world.

This paper develops a unified model of firms’ global R&D and production decisions, featur-ing talent-acquisition and market-access motives for offshore R&D, to study the determinants andwelfare implications of offshore R&D. Disciplined by estimates from a new dataset covering R&Dand production of firms from a large number of countries, the model can account for the ob-served level of offshore R&D. Quantitatively, offshore R&D brings about 3.3% welfare gains andamplifies the gains from openness by a factor of 1.3. Further experiments show that a country’sopenness to offshore R&D and offshore production have very different spillover effects to othercountries. Moreover, because of the interconnection between integration via trade, offshore R&D,and offshore production, incorporating offshore R&D makes a substantive difference when eval-uating other forms of integration. All these results point to the importance of offshore R&D inquantitative models of globalization.

This paper abstracts from some aspects of the reality that might prove useful for both mea-suring and theorizing about offshore R&D. For example, the model incorporates different tasks inbringing a product to consumers but has overlooked the role of sectors. Incorporating sectors andinput-output linkages can shed light on the role of sectoral comparative advantages and bring myanalyses into contact with the sector-based perspective of GVC measurements. Second, this papertakes the distribution of firm knowhow as given. Making knowhow accumulation an endogenouschoice can help understand the dynamic effects of integration via offshore R&D.

45

References

Alviarez, Vanessa, “Multinational production and comparative advantage,” Journal of Interna-tional Economics, 2019, 119, 1–54.

Antràs, Pol and Stephen Yeaple, “Multinational Firms and the Structure of International Trade,”Handbook of International Economics, 2014, 4, 55., Davin Chor, Thibault Fally, and Russell Hillberry, “Measuring the Upstreamness of Produc-tion and Trade Flows,” American Economic Review, 2012, 102 (3), 412–16.

Antras, Pol, Luis Garicano, and Esteban Rossi-Hansberg, “Offshoring in a Knowledge Econ-omy,” The Quarterly Journal of Economics, 2006, 121 (1), 31–77.

Argente, David, Douglas Hanley, Salome Baslandze, Sara Moreira et al., “Patents to Products:Innovation and Firm Performance,” Unpublished Manuscript, 2018.

Arkolakis, Costas, Natalia Ramondo, Andrés Rodríguez-Clare, and Stephen Yeaple, “Innova-tion and Production in the Global Economy,” American Economic Review, 2018, 108 (8), 2128–73.

Bilir, L Kamran and Eduardo Morales, “Innovation in the Global Firm,” Journal of Political Econ-omy, 2020, 128 (4), 1566–1625.

Bloom, Nicholas, Christos Genakos, Raffaella Sadun, and John Van Reenen, “ManagementPractices Across Firms and Countries,” The Academy of Management Perspectives, 2012, 26 (1),12–33.

Branstetter, Lee, Guangwei Li, and Francisco Veloso, “The Globalization of R&D: China, India,and the Rise of International Co-Invention,” Unpublished Manuscript, 2013.

Burstein, Ariel T and Alexander Monge-Naranjo, “Foreign Know-how, Firm Control, and theIncome of Developing Countries,” The Quarterly Journal of Economics, 2009, p. 149.

Cravino, Javier and Andrei A Levchenko, “Multinational Firms and International Business CycleTransmission,” The Quarterly Journal of Economics, 2017, 132 (2), 921–962.

de Gortari, Alonso, “Disentangling Global Value Chains,” Working Paper 25868, National Bureauof Economic Research May 2019.

Fan, Jingting, “Essays on the Welfare Implications of International Economic Integration.” PhDdissertation, University of Maryland 2017.

Garetto, Stefania, “Input Sourcing and Multinational Production,” American Economic Journal:Macroeconomics, 2013, 5 (2), 118–51., Lindsay Oldenski, and Natalia Ramondo, “Multinational Expansion in Time and Space,”Working Paper 25804, National Bureau of Economic Research May 2019.

Guadalupe, Maria, Olga Kuzmina, and Catherine Thomas, “Innovation and Foreign Owner-ship,” The American Economic Review, 2012, 102 (7), 3594–3627.

Haas, Martine R and Prithwiraj Raj Choudhury, “Collaborate to Innovate? Teams and PatentGeneration in a Global R&D Center,” Unpublished Manuscript, 2015.

Hanushek, Eric A and Ludger Woessmann, “Do Better Schools Lead to More Growth? CognitiveSkills, Economic Outcomes, and Causation,” Journal of Economic Growth, 2012, 17 (4), 267–321.

Head, Keith and John Ries, “Increasing returns versus national product differentiation as an ex-

46

planation for the pattern of US-Canada trade,” American Economic Review, 2001, 91 (4), 858–876.and Thierry Mayer, “Brands in motion: How frictions shape multinational production,” Amer-ican Economic Review, 2019, 109 (9), 3073–3124.

Helpman, Elhanan, “A Simple Theory of Ttrade with Multinational Corporations,” Journal of Po-litical Economy, 1984, 92 (3)., Marc J Melitz, and Stephen R Yeaple, “Export Versus FDI with Heterogeneous Firms,” Amer-ican Economic Review, 2004, 94 (1), 300–316.

Irarrazabal, Alfonso, Andreas Moxnes, and Luca David Opromolla, “The Margins of Multina-tional Production and the Role of Intrafirm Trade,” Journal of Political Economy, 2013, 121 (1),74–126.

Johnson, Robert C, “Measuring Global Value Chains,” Annual Review of Economics, 2018, 10, 207–236.and Guillermo Noguera, “Accounting for Intermediates: Production Sharing and Trade in

Value Added,” Journal of international Economics, 2012, 86 (2), 224–236.Keller, Wolfgang and Stephen Ross Yeaple, “The Gravity of Knowledge,” American Economic

Review, 2013, 103 (4), 1414–1444.Kerr, Sari Pekkala and William R Kerr, “Global Collaborative Patents,” The Economic Journal,

2018, 128 (612), F235–F272.Koopman, Robert, Zhi Wang, and Shang-Jin Wei, “Tracing Value-added and Double Counting

in Gross Exports,” American Economic Review, 2014, 104 (2), 459–94.Loecker, Jan De and Jan Eeckhout, “Global Market Power,” Working Paper 24768, National Bu-

reau of Economic Research June 2018.Markusen, James R, “Multinationals, Multi-Plant Economies, and the Gains from Trade,” Journal

of International Economics, 1984, 16 (3), 205–226.McGrattan, Ellen R and Edward C Prescott, “Openness, Technology Capital, and Development,”

Journal of Economic Theory, 2009, 144 (6), 2454–2476.Nocke, Volker and Stephen Yeaple, “An Assignment Theory of Foreign Direct Investment,” The

Review of Economic Studies, 2008, 75 (2), 529–557.Park, Walter G, “International patent protection: 1960–2005,” Research policy, 2008, 37 (4), 761–766.Ramondo, Natalia and Andrés Rodríguez-Clare, “Trade, Multinational Production, and the

Gains from Openness,” Journal of Political Economy, 2013, 121 (2), 273–322.Thursby, Marie and Jerry Thursby, Here or There?: A Survey of Factors in Multinational R&D

Location–Report to the Government-University-Industry Research Roundtable, National AcademiesPress, 2006.

Timmer, Marcel P, Abdul Azeez Erumban, Bart Los, Robert Stehrer, and Gaaitzen J De Vries,“Slicing up global value chains,” Journal of Economic Perspectives, 2014, 28 (2), 99–118.

Tintelnot, Felix, “Global Production with Export Platforms,” The Quarterly Journal of Economics,2016, p. qjw037.

Wang, Zi, “Headquarters Gravity: How Multinationals Shape International Trade,” UnpublishedManuscript, 2019.

47