Global Sourcing and Domestic Production Networks

This paper studies how firms’ offshoring decisions shape a country’s domestic production net- works. We develop a model in which heterogeneous firms source inputs from multiple industries located in different domestic regions and foreign countries. Input sourcing entails communication with suppliers, which is endogenously increasing in the differentiation of inputs. The model predicts that firms are less likely to source differentiated inputs, especially from distant domestic and foreign suppliers, due to costly communication. Triggered by foreign countries’ export supply shocks, firms start offshoring inputs from foreign suppliers, which displace the less productive domestic suppliers in the same industry (the direct displacement effect). The resulting decline in marginal costs induces firms to start sourcing from the more productive and distant domestic suppliers within industries (the within-industry restructuring effect), but possibly also from nearby suppliers that produce inputs that are more differentiated than those supplied by existing suppliers (the industry composition effect). The net effect of offshoring on a firm’s domestic production networks depends on the relative strength of the three effects, which we verify using data for 4.5 million buyer-seller links in Japan. Based on a firm-level instrument, we find that after offshoring, firms are less likely to drop suppliers on average, but more so for the larger ones. They tend to add nearby suppliers producing differentiated inputs. These results suggest that firms offshoring may increase the spatial concentration of domestic production networks.


Introduction
Substantial declines in trade barriers and advances in information, communication and transportation technologies have encouraged more …rms to source inputs from far-away suppliers. A growing body of literature studies both the causes and consequences of increasing global production fragmentation. 1 The focus of the literature has been the direct e¤ects of global sourcing on the industry or …rm that imports intermediate inputs, despite the fact that an economy is an interlinked web of production units, each using inputs from its suppliers to produce goods and services that are sold further downstream. Indeed, recent research has shown the signi…cance of considering production networks for a wide range of economic topics, such as the propagation and ampli…cation of …rm-level shocks to large business-cycle ‡uctuations (Acemoglu et al., 2012;Carvalho and Gabaix, 2013); knowledge spillover (Javorcik, 2004); the aggregate e¤ects of resource misallocation (Jones, 2011 and; and the gains from trade (Costinot and Rodriguez-Clare, 2014; Caliendo and Parro, 2015). How global trade shapes the production networks in a country is an important yet relatively under-explored topic. This paper studies from both theoretical and empirical perspectives how …rms' sourcing of intermediate inputs from foreign suppliers, which we refer to as o¤shoring, reshapes a country's domestic buyer-supplier networks. Speci…cally, we examine the e¤ects of o¤shoring that is triggered by foreign cost shocks on …rms'choices of domestic input suppliers. To guide our empirical analysis, we extend the global sourcing model by Antràs, Fort, and Tintelnot (2017, henceforth AFT) to consider multiple domestic source regions, various input industries that di¤er in product di¤erentiation, and face-to-face communication between heterogeneous buyers and input suppliers. It builds on the premise that trade is more costly over longer distance, and especially so when the success of input production depends on the intensity of communication between buyers and suppliers. Similar to Bernard, Moxnes, and Saito (2016, henceforth BMS), our model features heterogeneous buyers and sellers engaged in costly domestic trade; but we additionally analyze …rms'input sourcing from both domestic and foreign suppliers in di¤erent input industries. We then empirically examine the sourcing is stronger for the more di¤erentiated inputs, as measured by the inverse of the elasticity of substitution between input varieties. Hence, …rms are less likely to source di¤erentiated inputs from the more distant regions or from foreign countries. Only the relatively more productive …rms will outsource di¤erentiated inputs, with the most productive ones o¤shoring them.
Besides portraying the spatial and sectoral patterns of …rms'global sourcing, we use the network data to examine the model predictions about the e¤ect of o¤shoring on …rms'choices of domestic suppliers. To establish the causal link between …rms'o¤shoring and the pattern of domestic sourcing, we construct a …rm-level instrument using information on buyers'initial patterns of domestic sourcing across industries and the corresponding foreign countries'export supply shocks. The idea is that conditional on a …rm's sourcing inputs from a domestic input industry, the incremental …xed costs needed for o¤shoring inputs in the same industry are lower. When positive export supply shocks, due either to reduced trade costs or increased productivity of Japan's trade partners, hit an industry, those that are already sourcing inputs in the same industry should …nd o¤shoring relatively more attractive.
The two-stage least squares estimates show that o¤shoring induces …rms to add and drop larger domestic suppliers simultaneously, while adding the more proximate ones. While the addition of the larger suppliers is consistent with the within-industry restructuring e¤ect of o¤shoring, both dropping of the larger suppliers and adding of the closer suppliers need to be explained by the other two e¤ects. The dropping of the larger suppliers imply a su¢ ciently strong direct displacement e¤ect, which is stronger in the less di¤erentiated industries; while the adding of the more proximate suppliers is due to the industry composition e¤ect, which implies the addition of suppliers from di¤erentiated industries from which the buyers did not source inputs previously. Consistently, we …nd that newly o¤shoring …rms are more likely to start sourcing from the more di¤erentiated input industries. These results along with the higher likelihood of adding the more proximate suppliers suggest a strong industry composition e¤ect, and o¤er an explanation for why o¤shoring can reduce the average distance between buyers and suppliers in production networks. The documented patterns of supplier reorganization also have implications for analyzing how o¤shoring a¤ects aggregate productivity. 6 Our paper relates to several strands of literature. First, it contributes to the growing literature on production networks and international trade. The study by Atalay, et al. (2011) is one of the …rst in the literature to theoretically and empirically describe …rms'production networks in a country (the U.S.). Ober…eld (2017) develops a model to study the endogenous formation of …rms' production networks and its impact on aggregate productivity. Thanks to the recently available buyer-seller linked data, there is a burgeoning literature studying the pattern and dynamics of domestic production networks. 7 Notably, BMS use the same Japanese buyer-seller linked data to highlight the negative assortativity of buyer-supplier links. Based on the extension of the highspeed train line in Japan as an exogenous shock, the authors …nd that …rms near newly built train stations tend to increase sourcing from more domestic locations and thereby experience an increase in measured productivity. Using the same data and exploiting the Great East Japan Earthquake of 2011 as an exogenous shock, Carvalho et al. (2017) quantify the propagation of shocks through the domestic input-output linkages. 8 Tintelnot et al. (2017) develop and estimate a model of …rm-to-…rm domestic trade, foreign trade, and endogenous network formation. Di¤erent from all these studies, our paper focuses on characterizing both the industrial and geographic patterns of domestic production networks. Importantly, we are the …rst to incorporate both domestic and foreign sourcing in a model to empirically examine how …rms' o¤shoring shapes the domestic production networks across sectors and space. This paper also contributes to the growing cluster of work on networks in international trade. 9 Recent research seeks to study the micro foundation of the dynamics and patterns of …rms'sorting and matching in international trade networks (e.g., Chaney, 2014;Eaton et al., 2014;Carballo, Ottaviano, and Volpe Martincus, 2016; Bernard, Moxnes and Ulltveit-Moe, 2017; Sugita, Teshima, and Seira, 2017). 10 In particular, Bernard, Moxnes and Ulltveit-Moe (2017) build a model that feaprocurement share form 25% in 2011 to 70% in 2021. A new overseas procurement base will be built in India, in addition to their existing bases in Thailand and China. As part of this o¤shoring plan, the company would need to reorganize the procurement relationships with the existing domestic suppliers. 7 Using U.S. buyer-seller linked data and a structural model of …rms' network formation, Lim (2017) studies the macroeconomic implications of the propagation of …rm-level demand and supply shocks through the production networks. 8 They show that external shocks on downstream …rms a¤ect not only the directly linked upstream …rms, but also …rms that are two or three degrees away from the a¤ected …rms. 9 The literature dates back to the seminal work by Rauch (1999) and Rauch and Trindade (2002), who show that colonial ties, common languages, and the stock of immigrants between two countries are positively related to bilateral trade, especially for di¤erentiated products. The authors relate these …ndings to the importance of networks, information and search frictions in trade. See Chaney (2016) for a literature review.
1 0 Using importer-exporter matched data from Colombia, Eaton et al. (2014) structurally estimate the e¤ects of 6 tures two-sided heterogeneity and uncover in Norwegian importer-exporter linked data the negative assortative matching of trade partners. 11 Chaney (2014) proposes theoretically that a country's aggregate export dynamics are tightly linked to …rms'penetration into new foreign markets, through establishing new contacts and expanding existing trade relationships. Di¤erent from this literature, which focuses primarily on the patterns of importer-exporter matches, our paper focuses instead on …rm-to-…rm relationships in the domestic economy, and investigates the impact of …rms'o¤shoring on the evolution of the domestic segment of global value chains.
Our work also contributes to the literature on the interplay between international trade and the non-e¢ ciency aspect of …rm performance. In particular, Holmes and Stevens (2014) …nd in the U.S. manufacturing …rm census data that small plants specialize in making specialty goods sold to nearby customers, while large plants specialize in mass production of standardized goods shipped to distant markets. Motivated by these facts, the authors structurally examine …rms'heterogeneous responses to import shocks from China, which cannot be explained by a standard heterogeneous-…rm model. 12 Relatedly, we show paradoxically that large domestic suppliers are more likely to be dropped, while proximate suppliers are more likely to be added, by the newly o¤shoring downstream …rms. 13 In this regard, this paper shares similar key messages with Jensen and Kletzer (2005), who study the tradability of tasks, and Keller and Yeaple (2013), who examine the ways multinationals transfer knowledge to their overseas a¢ liates based on sector characteristics. In particular, we show using …rm-to-…rm linked data that the elasticity of trade costs with respect to distance varies across industries, and thus a¤ects …rms'reorganization of domestic production networks upon o¤shoring. In a sample with mostly one-to-one matches, the authors …nd evidence of positive assortative matching in trade. 1 1 Similar to Carballo, Ottaviano, and Volpe Martincus (2016), the authors also highlight adjustments on the buyer margin as an important channel through which trade responds to policy shocks. 1 2 In particular, the authors …nd that large rather than small plants experience the largest contraction in sales in response to the import shocks, in contrast to the predictions of a standard heterogeneous-…rm model. 1 3 Another dimension of …rm performance is product quality, which has been studied by a large and growing literature, such as Khandelwal (2010) (Kasahara and Rodrigue, 2008; Hijzen, Inui,   and Todo, 2010; AFT, 2017). 15 It contributes to the economic geography literature by showing that o¤shoring can be a source of industry coagglomeration, as generic input suppliers, which are on average located farther away, are the ones that tend to be displaced by foreign suppliers, while di¤erentiated input suppliers, which are on average located nearby, tend to be added as new suppliers by o¤shoring …rms. Contributing to the literature on the (direct) e¤ects of o¤shoring, our results about the patterns of supplier adding and dropping highlight a previously omitted channel through which o¤shoring can a¤ect an economy's labor market outcomes and aggregate productivity.
The paper proceeds as follows. Section 2 discusses the data sources. Section 3 presents several stylized facts that motivate a theoretical model, which is introduced in Section 4. Section 5 discusses our empirical design and …ndings. The …nal section concludes this paper.

Data
Our data come from two sources. The network data for 2005 and 2010 come from Tokyo Shoko Research Ltd. (TSR), a private credit reporting agency. Firms provide information to TSR in order to obtain credit scores for loans. The TSR data contain basic …rm-level balance sheet information, such as employment, sales, location, main (4-digit) industry (up to 3), founding year, number of 1 4 Ellison and Glaeser (1997) propose sectoral measures of the degree of industry agglomeration and coagglomeration, and …nd evidence of coagglomeration in industry pairs with strong upstream-downstream relationships. Rosenthal and Strange (2001) and Ellison, Glaeser, and Kerr (2010) empirically identify causes of agglomeration and coagglomeration, such as knowledge spillovers, input sharing, product shipping costs, labor market pooling, and natural advantage. Rosenthal and Strange (2001) …nd that labor market pooling has the most robust e¤ect, while Ellison, Glaeser, and Kerr (2010) …nd evidence that input-output linkages are particularly important. Using Japanese buyer-seller linked data, Nakajima, Saito, and Uesugi (2012) …nd evidence that intensity of intra-industry transactions increases industry agglomeration. On the trends of industry agglomeration, Dumais et al. (2002) investigate dynamics of geographic concentration of U.S. manufacturing industries. They …nd that although the trend of industry agglomeration varies with industries, their average agglomeration levels have declined slightly in recent decades. Behrens and Bougna (2015) also observe a recent decline in the agglomeration of manufacturing plants in Canada. 1 5 Ebenstein et al. (2014) and Hummels et al. (2014) examine the e¤ect of o¤shoring on workers'wages using U.S. and Danish data, respectively. Kasahara and Rodrigue (2008) use Chilean manufacturing plant-level data and …nd that …rms' importing of intermediates improves productivity. Hijzen, Inui, and Todo (2010) …nd a positive impact of o¤shoring on Japanese …rms'productivity. AFT build a multi-country sourcing model to study both theoretically and empirically …rms'selection into importing and the resulting cost e¤ects and complementarities between sourcing from di¤erent countries. 8 establishments, of over 800,000 …rms in Japan. 16 Crucially, it also provides information on …rmto-…rm relationships. Each …rm surveyed by the TSR was asked to report the names of its top 24 suppliers, top 24 customers, and 3 main shareholders. To avoid the "top 24" cuto¤ from limiting the sample coverage of the production network, we use a two-way matching method to maximize the number of links, using information reported by a buyer about its sellers and vice versa. Since a relationship with a buyer or seller can be reported by either end of a relationship, the number of buyers (sellers) of a seller (buyer) can be much greater than 24. In fact, the top seller in our constructed network data in Japan has over 11,000 buyers in 2010, while the top buyer has close to 8,000 suppliers. The distribution of the buyer-supplier links is very skewed, with most of the …rms having substantially fewer buyers and sellers (more below). Distance between any pair of buyers and sellers is measured using the addresses reported by the …rms, which we geocode. 17 We complement the TSR data with the Basic Survey of Japanese Business Structure and Activities (BSJBSA), conducted annually by the country's Ministry of Economy, Trade and Industry (METI). The BSJBSA data cover all …rms that have over 50 employees or 30 million yen of paid-in capital in the country's manufacturing, mining, wholesale and retail, and several service sectors.
Firms' responses to the survey are mandatory. The survey data contain detailed information on …rms'business activities, such as their main industry (3-digit), number of employees, sales, capital (which is required to compute a …rm's total factor productivity), purchases of inputs and materials, exports and imports by continent (e.g., Asia, Europe, etc.). 18 The data set covers 22,939 and 24,892 …rms for 2005 and 2010, respectively. We merge the two data sets using …rms'names, addresses, and telephone numbers. The merged data contain over 800,000 buyer-supplier pairs. 19 In the regression analysis, we focus on the subsample that has manufacturing …rms on the buyer side of a relationship. 1 6 The surveys were conducted in 2006 and 2011, respectively. We use both TSR Company Information Database and TSR Company Linkage Database in this paper. According to Carvalho et al. (2017), the TSR data cover more than half of all …rms in Japan. According to BMS, the TSR sample covers almost all …rms with over 4 employees in Japan. 1 7 We use the geocoding service from the Center for Spatial Information Science at the University of Tokyo to …rst identify the latitude and longitude of each address, and then compute the distance between any pair of coordinates. 1 8 The data set, however, does not provide information on …rms'imports by sector. 1 9 About half of the observations of the balanced TSR sample have buyers that can be merged to …rms included in the manufacturing survey. See Table A3 in the appendix about the summary statistics of the key variables from the BSJBSA data. Importers'average imports-to-intermediate ratio, increases from 18% to 21% from 2005 to 2010. Asia is a very important input source for Japanese importers-among importers, the average share of imports from Asia is over 80% in both 2005 and 2010. 9 3 Descriptive Evidence

Domestic Production Networks
We …rst describe several key patterns observed in our network data. Table 1 reports the summary statistics on the buyer-supplier links. Panel A reports that the number of links, based on the TSR sample, is about 3.6 millions in 2005 and 4.5 millions in 2010. The average number of sellers for a buyer increased from 4.9 in 2005 to 5.5 in 2010, while the median increased from 2 to 3. The large di¤erence between the mean and the median numbers of sellers per buyer suggests a highly skewed distribution of buyer-supplier links (i.e., a small number of large buyers have substantially more sellers than other buyers). 20 The increases in the average and median numbers of buyer-supplier links since 2005 suggest that the production network in Japan is getting denser. Since the rise in the density of the network may be due to …rms' more comprehensive self-reporting of sellers, we use a regression sample that includes only buyers and suppliers that operated in both 2005 and 2010, to mitigate this potential measurement issue. 21 Panel B reports the summary statistics of the number of links in the regression sample built from merging the BSJBSA …rm sample with the TSR network data. Since BSJBSA imposes sampling thresholds based on …rms' employment and capital, the mean and the median numbers of sellers linked to a buyer in our regression sample are larger (25 and 10, respectively, for the year 2005) than those in the network data. 22 Tables A1 and A2 in the appendix report more detailed statistics by buyers'main industry. Table 2 reports the summary statistics on the numbers of sellers and domestic regions (47 prefectures) from which di¤erent types of buyers, based on their import statuses, sourced inputs.
2 0 When we plot the log number of sellers of a buyer against the fraction of buyers having at least that many sellers ( Figure A1 in the appendix), we …nd a power-law distribution, as highlighted by BMS. 2 1 The cost of using a data set with balanced numbers of buyers and sellers in both years is that we cannot study the entries and exits in the network. 2 2 One may be concerned about the selection biases arising from the exclusion of small …rms in our regression sample. Three remarks are in order. First, if the goal of the study is to evaluate the e¤ects of …rms' o¤shoring on their choices of domestic suppliers, the focus on large …rms should be …ne as large …rms are more likely to engage in o¤shoring, which entails high …xed costs (see AFT for the structural estimate of those …xed costs). Second, if there is any e¤ect of o¤shoring on …rms'performance and therefore their choices of domestic suppliers, omitting small …rms, which tend to be non-importers, in our sample will go against us from …nding any e¤ect. It is because the variation in …rms'participation in o¤shoring and the associated e¤ects would have been even larger if small …rms were included in the sample. Third, even though the fraction of …rms that have at least n links is naturally larger based on the regression sample, the power-law distribution of the number of sellers per buyer is preserved (see Figure A1 in the appendix). The slope of the relation based on the regression sample is almost identical to that derived from the original TSR sample.
In 2005, there are altogether 13,784 manufacturing buyers in the regression sample. Of these buyers, 7% did not import in 2005 but started importing by 2010, while 74% continued to be non-importers by 2010. Firms that imported in both 2005 and 2010 accounted for about 13% of the sample. 23 These continuing importers sourced inputs from more domestic sellers and prefectures than new importers and non-importers. They procured inputs from 48.5 domestic sellers in 2005 on average, with the median equal to 16. The mean and median numbers of prefectures from which existing importers procure inputs are 7.49 and 5, respectively. For new importers, while their numbers of sellers and source prefectures are on average smaller than those of continuing importers, they are larger than those of continuing non-importers.  Figure 1 shows that the number of buyer-seller links is negatively correlated with the distance between the pair, and about half of the connections are observed within a 25 km radius of buyers. This negative correlation appears to increase in magnitude since 2005. Figures 2 and 3 reveal the relationship between a buyer's sales and its scope of domestic sourcing. Figure 2 shows a positive correlation between buyers' sales and the number of connected domestic suppliers, while Figure 3 depicts a positive correlation between their sales and the number of prefectures from which they source inputs. We also …nd that within the same 4-digit industry and prefecture, the more productive buyers use more suppliers, and distant suppliers tend to be more productive (see Table A4 in the appendix for the regression results). These results altogether demonstrate the importance of incorporating trade frictions that are increasing in distance, along with two-sided heterogeneity across buyers and sellers in the model, similar to BMS.

Firms'Post-O¤shoring Outcomes
Let us now present some preliminary empirical results about the correlation between a …rm's o¤shoring (importing) status and its post-o¤shoring performance. We estimate the following spec-i…cation using a simple …xed e¤ects model: where is an operator that takes the …rst di¤erence of the variable y i between 2005 and 2010, while y i represents buyer i's log sales or various measures of the scope of domestic sourcing, including log numbers of domestic suppliers, domestic industries, and domestic regions respectively from which buyer i sources inputs. We also examine how o¤shoring changes the average distance between a buyer and its domestic suppliers. We use three measures of the change in a buyer's average distance from its suppliers, represented by dist i;t . The …rst one is the Davis-Haltiwanger-Schuh (1998) growth rate of distance, (dist i;10 dist i;05 ) = 1 2 (dist i;10 + dist i;05 ), which is by construction bounded between -2 and 2 to reduce the impact of outliers. The second measure is the log di¤erence in the average distance. The third measure considers the di¤erence between the average distance of the newly added suppliers and that of the dropped suppliers, dist add The variable of interest, imp i , represents the change in …rm i's import status, which equals 1 if buyer i did not import in 2005 but started to import in 2010, 0 otherwise. 24 We include buyer's (4-digit) industry and region (prefecture) …xed e¤ects (F E s and F E r ) to control for any unobserved determinants of …rm outcomes (e.g., …rms in certain prefectures are more likely to source inputs due to a high geographic concentration of suppliers). We always control for buyer i's 2005 log total factor productivity, T F P i , as it is well documented in the literature that more productive …rms are more likely to import intermediate inputs (e.g., Amiti, Itskhoki, and Konings, 2014). Table 3 reports the estimates of Equation (1). The regression sample includes manufacturing buyers only, while the construction of a buyer's performance measure uses information of its linked suppliers in both manufacturing and non-manufacturing industries. 25 Column 1 reports a positive and signi…cant correlation between the change in the …rm's import status and the change in its sales. From columns 2 to 4, we …nd a positive and signi…cant correlation between the change in the …rm's import status and its scope of domestic sourcing, as measured by the (log) numbers of suppliers, industries, or regions from which the …rm sources its intermediate inputs.
In columns 5 and 6, we …nd a signi…cant and negative correlation between a …rm's o¤shoring 2 4 Recent research reveals that many exporters only export for a year and then drop out from exporting (e.g., Blum, Claro, and Horstmann, 2013). To address the issues of occasional importing, we conduct robustness checks by de…ning a new importer as one that imported for two consecutive years (2010 and 2011), and a non-importer as one that did not import for three consecutive years (2003)(2004)(2005). The main results remain robust and are available upon request. participation and the average distance between the …rm and its domestic suppliers. In column 7, we …nd that after a …rm starts o¤shoring, the distance from the newly added suppliers relative to that of the dropped suppliers tends to drop. 26 Figures 4 and 5, which show the links of added and dropped domestic suppliers of electronics …rms that started o¤shoring inputs since 2005, portray a pattern that suggests an increased geographic concentration of domestic production networks. Without any causal implications, the results in Table 3 show insightful correlation between o¤shoring and …rms'domestic sourcing behavior. In Section 5, we will propose a …rm-level instrument to gauge the causal e¤ect of o¤shoring on …rms'adding and dropping of suppliers.
While the positive correlation between a …rm's import participation and the scope of domestic sourcing is consistent with the main …ndings in AFT, the negative correlation between o¤shoring and the average buyer-supplier distance cannot be readily rationalized by their model that only considers a single input industry. We therefore develop our own model in the following section.

A Model of Firms'Global Sourcing
Motivated by the suggestive evidence above, we develop a model that features heterogeneous …rms' sourcing of intermediate inputs from suppliers located in di¤erent domestic and foreign regions.

Set-up
We consider a representative industry and build an industry equilibrium model that features global sourcing (domestic sourcing and o¤shoring). Our model extends AFT to study the pattern of global sourcing and the e¤ect of o¤shoring on …rms'domestic production networks. Similar to BMS, our model considers input suppliers located in multiple domestic regions. Unlike their single-industry models, however, we consider multiple input industries that di¤er in the degree of product di¤erentiation. We investigate how the di¤erentiation of inputs changes …rms' incentives to outsource inputs and how o¤shoring a¤ects …rms'post-o¤shoring relationships with individual suppliers. We also introduce in the model buyers'communication with sellers in an e¤ort to enhance input quality to show that the elasticity of trade costs with respect to distance can endogenously increase with the di¤erentiation of inputs. We …rst characterize …rms'spatial and sectoral equilibrium patterns of global sourcing, before examining the e¤ect of a reduction in foreign input costs on buyers'choices of domestic suppliers.

Demand
Consider an industry facing only domestic demand. 27 The industry has a continuum set N of exogenously-given …nal-good producers of horizontally di¤erentiated products. Consumers have a common love-of-variety utility function that features constant elasticity of substitution (CES), denoted by > 1. Each …rm i faces its own demand: is the price index and E is the total expenditure on the goods. Since …nal-good producers are the buyers of intermediate inputs in the model, they will be referred to as buyers while input suppliers will often be referred to as sellers.

Final Goods Production
Final goods are produced with inputs from S di¤erent industries, which di¤er from each other in the degree of product di¤erentiation. Production of …nal goods involves two stages. The …rst stage is to make S composite inputs, each with a unit mass of di¤erentiated input varieties, using the following CES production function: wherex is denotes the quantity of composite input s 2 f1; ; Sg that is produced and used by …rm i for …nal-good production, while x is (j) denotes the quantity of variety j of input industry s.
The parameter s > 1, which is the elasticity of substitution between di¤erent input varieties in the production of composite input s, is our (inverse) measure of input di¤erentiation. Intuitively, an input is di¤erentiated if it has to be tailored to the speci…c needs of a buyer and is therefore di¢ cult to be substituted with other varieties produced by other suppliers. As such, input varieties that are less substitutable are considered more di¤erentiated. We order input industries such that a higher index s indicates a higher degree of product di¤erentiation (i.e., 1 > 2 > > S ).
The second stage of the …nal-good production is to assemble S composite inputs into …nal goods.
The assembly technology of buyer i takes the Cobb-Douglas form: where s is the cost share of each input industry in producing …nal goods while ' i is buyer i's core productivity.
Each variety j of every input industry s can be insourced or sourced from a supplier located in Any input or …nal-good producer independently draws its input-production productivity z from a Fréchet distribution with a cumulative distribution function de…ned over (0; 1) by where T sr > 0 is positively related with the likelihood of a high-productivity draw while s > 1 governs the inverse variability of the draws. A smaller s implies a larger variation of productivity across …rms within the sector. The location parameter T sr can vary across …nal-good producers (r = 0) and input producers (r 6 = 0), as well as across regions. 29 An input supplier with productivity z has a unit cost of production of w r c s =z, where w r is a region-speci…c cost parameter such as the wage rate while c s is the cost parameter that is speci…c to the input industry.

Trade Costs, Buyer-Seller Communication, and Input Quality
Outsourcing requires a buyer to incur two types of …xed costs. The …rst type is the cost to make inputs "outsourceable" (e.g., to codify the design of inputs). Speci…cally, for each input industry in which a buyer outsources intermediate inputs, it has to incur a …xed cost of f , which is common across input industries. The second type of …xed costs are those related to search in a region for the lowest-cost sellers of individual input varieties. Borrowing the insight from AFT, we assume that for every region a buyer searches for input suppliers, it incurs an industry-speci…c …xed cost of f s . Costly search implies that …rms will not source inputs from all regions. Let is denote the set of regions from which …rm i sources inputs in industry s. The set is may be a proper subset of f1; ; M; M + 1; ; M + M g. We assume that no …xed cost is required for insourcing, so that a buyer will always insource a fraction of varieties even in the industries that it outsources inputs.
There are also standard iceberg transport costs for domestic and foreign trade of inputs. They take the form t s (d) 1 and t s (0) = 1, where t s is an industry-speci…c increasing function of the distance d between the buyer's region and a seller's region.
The transport cost, however, is not the only trade cost that increases with the distance between buyers and sellers. Buyers need to communicate with sellers to make sure that they receive what they want. The cost of face-to-face communication naturally increases with distance, and its bene…t clearly depends on the di¤erentiation of the inputs that are traded. 30 Consider a misunderstanding between a pair of buyer and seller about the speci…cation of a product (e.g., size, shape, and color).
Low-quality parts and components may reduce the quality of …nal products at the minimum, and can jeopardize the entire production process in the extreme situation. 31 Based on the presumption that the failure of delivering high-quality inputs often arises from miscommunication or misunderstanding between buyers and sellers, we assume that a buyer can reduce the probability of failure by engaging in face-to-face communication more vigorously. 32 More speci…cally, we assume that for each input variety j 2 [0; 1] in industry s, a seller's products meet the buyer's expected standard with probability q, and fail to meet the standard with probability 1 q. We further assume that in the latter case, all inputs produced by that seller are useless for the buyer. Buyers, however, can a¤ect q by engaging in communication with individual input suppliers, which raises the unit cost of shipped inputs by a multiple of e m(d)q , where m is an increasing function of the distance between the buyer and a seller. 33 The marginal communication cost rises with the distance (i.e., face-to-face communication with distant sellers is more costly).
In this model, we assume for simplicity that buyers have all the bargaining power against input suppliers, so that the price of an input equals its unit cost. 34 Given a productivity distribution f' i g i2N , each buyer i makes a sequence of decisions as follows: 1. Buyer i as well as each input supplier draws its productivity for input production. Buyer i knows its own productivities for input production for all j 2 [0; 1] in every input industry s = 1; ; S.

2.
In every input industry s, buyer i chooses whether to outsource or not, and pays f for every industry that it has chosen to outsource. In addition, for each industry s that it has chosen to outsource, it selects a set of regions that it searches for input suppliers, and pays f s for every such region. 5. Buyer i optimally sets its …nal-good price, which will be a constant mark-up over its marginal cost. 3 3 By specifying the communication as a variable cost rather than a …xed cost, we assume that the intensity of communication is increasing in the value of transaction. The …nding of a positive correlation between the total value of outsourcing and the number of business travels at the industry level for the U.S. o¤er indirect evidence for that assumption (Liu, Scholnick, and Finn, 2017). 3 4 Introducing explicit negotiation between buyers and sellers would not change the results qualitatively.

Optimal Communication Intensity
We now derive each …rm i's optimal communication intensity, characterized by the probability q = q isr that …rm i receives high-quality inputs, taking its set of source regions, f is g S s=1 , as given.
For a given set of suppliers in the regions in is and hence a given set of prices for the input varieties of industry s, buyer i chooses q isr to minimize the e¤ective unit cost of the composite input s. Let G isr denote the probability distribution of the price of inputs sourced from region r.
Also let I isr denote the set of inputs sourced from region r and (I isr ) its measure. Due to the law of large numbers, the mass (1 q isr ) (I isr ) of the input varieties sourced from region r 2 is is useless, while the prices of remaining q isr (I isr ) of input varieties are distributed according to the distribution of G isr . There is no such loss for insourced varieties.
Firm i optimally selects how much it purchases from each seller, given the risk of receiving useless inputs with probability 1 q isr . As shown in Appendix A, the resulting unit cost for the composite input s, denoted byc is , re ‡ects this risk: Note that for r 2 is , unit cost p is multiplied by q To alleviate the cost of receiving low-quality inputs, buyer i chooses q isr for each r 2 is to minimize q s 1 s isr p, which can be written as where d ir denotes the distance between buyer i and region r. It can be readily veri…ed that the cost-minimizing q isr is given by The communication intensity (or the probability of receiving high-quality inputs) and hence the communication costs decrease with s and d ir . Buyers have more incentive to enhance the communication with sellers of the more di¤erentiated inputs, since failing to obtain high-quality inputs is more costly due to a lower substitutability between input varieties. The communication incentive diminishes with the distance to the supplier because face-to-face communication, by assumption, is more costly over longer distance.

Optimal Sourcing Strategies
Let us turn to the stage in which each buyer i selects a seller for each input variety of industry s, taking the set of source regions as given. We will then solve backward for the optimal set of source regions.
The price of inputs …rm i buys from a seller, inclusive of trade costs (i.e., transport costs and communication costs), varies with the seller's productivity z and the distance to the seller's location d ir . In the case of insourcing, the price, or the unit cost, of an input variety is p = z 1 w 0 c s . For an input variety sourced from region r, it equals Note that all the price variations within the source regions come from the di¤erences in sellers' productivities.
Thus, as shown in Appendix A, we can apply the results of Eaton and Kortum (2002) to obtain buyer i's sourcing pattern and its costs of …nal-good production as follows.
The share of input varieties in industry s procured from region r is isr = is , with the sourcing potential isr given by and the sourcing capability by is is0 + P r2 is isr . Notice that in the absence of the communication channel, the trade elasticity is only s , as in Eaton and Kortum (2002) and many of its variants. Communication raises the cost of trade, making the trade elasticity also depend on s .
The …rm-speci…c unit cost of the composite input s, which is given by (3), can then be rewritten as where s s+1 s s 1 1 s , with (x) = R 1 0 t x 1 e t dt being the gamma function, and s < 1 + s 19 is assumed to hold.
We can now express the pro…t function for buyer i, still taking the optimal set of source regions as given. As shown in Appendix A, for a given cost pro…le fc is g S s=1 , …rm i's unit cost of …nal-goods production can be expressed as It immediately follows that …rm i's pro…ts can be expressed as where is takes 1 if buyer i outsources some inputs in industry s, and 0 if it insources all input varieties in industry s, and The pro…t function (8) conveys a lot of information about a …rm's optimal sourcing. Outsourcing input varieties in any industry s entails a …xed cost of f , while adding a new region r for sourcing inputs in industry s comes with an additional …xed cost f s . But they confer a bene…t of lowering the marginal cost of production, due to the expansion of the supplier set (i.e., an increase in is ).
Buyer i makes an optimal choice of the source regions, described by f is g S s=1 , based on balancing these costs and bene…ts. 35 There is no closed-form solution to the …rm's optimal choices of outsourcing and source regions.
However, we can still describe the buyer's optimal sourcing strategy through the …rst-order approximation of changes in i (' i ) in (8). The increment of i (' i ) when …rm i adds a region, say r 1 , to is 1 = , for some industry, say s 1 , can be approximated as is denotes …rm i's operating pro…ts. Whereas the increment of i (' i ) when it outsources inputs of industry s 1 at all can be approximated as where 6 = ;. A region is more likely to be added if isr is greater, which is in turn the case if (i) n s is larger, (ii) T sr is larger, (iii) w r is smaller, or (iv) d ir is smaller.
In equilibrium, inputs of industry s are outsourced if and only if (11) is nonnegative. In principle, buyer i chooses its source regions for each outsourced industry s by selecting the regions in a descending order from the region with the largest isr to the region with the smallest one as long as adding a region gives the buyer a net bene…t. However, such monotonicity of adding source regions may not always hold if s ( 1) < s , which AFT call the substitutes case. Appendix A shows some further details of the industry equilibrium, including its existence and uniqueness.

Global Sourcing
Having derived the industry equilibrium, we now discuss some features of global sourcing. We begin with the relationship between global sourcing and the productivity of buyers and sellers.
that the marginal bene…t of expanding the search increases with buyer i's core productivity ' i . The nesting property-the set of source regions weakly expands with the buyer's core productivity-is also obtained in what AFT call the complements case (i.e., when s ( 1) > s ). 36 Turning to the seller's productivity, our model predicts a negative correlation between the buyer-seller distance and the seller's productivity, which is similar to a …nding of BMS. It follows from (4) that the e¤ective price of inputs sourced from region r can be written as The e¤ective price of inputs sourced from a region increases with its distance from the buyer due to the increasing trade costs, arising from a smaller chance of receiving high-quality inputs and greater transport costs, while the distributions of the e¤ective price of the inputs outsourced are common across source regions as in Eaton and Kortum (2002). Consequently, inputs supplied from farther regions tend to be produced by more e¢ cient …rms than those in closer regions. Interpreting these results from the viewpoint of domestic versus foreign sourcing, we show that buyers with higher productivity tend to source from foreign suppliers and that foreign trade partners tend to be more cost-e¤ective than the domestic ones.
Let us turn to the examination of how buyers' sourcing strategies depend on the degree of input di¤erentiation. First, we show that the likelihood of outsourcing is negatively related to input di¤erentiation. It follows from (5) that the ratio of region r 1 's sourcing potential to …rm i's insourcing potential can be expressed as It can be readily shown that isr 1 = is0 is increasing in s . Since buyers choose a higher intensity of communication for the more di¤erentiated inputs, insourcing is relatively more appealing to them for such inputs because they need not engage in costly communication in the case of insourcing.
Once a buyer chooses to outsource some input varieties, it will then choose the optimal set of source regions. We show next that the negative correlation between distance and the sourcing potential is greater for the more di¤erentiated inputs. To compare the sourcing potential of region r 1 with that of another region r 2 , where d ir 1 > d ir 2 , we obtain from (5) the ratio of the sourcing potentials as isr 1 The multiplicative terms that involve distance, i.e., s s s 1 , are less than 1, which implies that the sourcing potential of the farther region r 1 tends to be smaller than that of the closer region r 2 . Moreover, isr 1 = isr 2 is smaller, the greater is the input di¤erentiation (i.e., the smaller is s ). Thus, we have shown that the more di¤erentiated inputs are more likely to be completely insourced and that distance matters more for the di¤erentiated inputs in …rms' 22 outsourcing decisions.

Proposition 1
The share of input varieties insourced and the share of input varieties sourced from closer regions are both greater for the more di¤ erentiated inputs.

Reduction in Foreign Input Costs and Restructuring of Production Networks
We now examine how a reduction in foreign input costs a¤ects …rms'o¤shoring decisions and their domestic sourcing strategies. We consider any changes that increase isr for some foreign region r 2 fM + 1; ; M + M g, including a fall in w r and an increase in T sr .
An increase in isr makes region r more attractive than before for all buyers. Consider the case in which an increase in isr induces some buyers to start sourcing inputs from region r .
Their individual sourcing capabilities, is , increase as a result, leading to lower marginal costs of production. The buyers that have been outsourcing some inputs from region r even before a reduction in foreign input costs also enjoy lower marginal costs, while those that do not source any inputs from region r experience no change in their marginal costs.
As the costs of o¤shoring from region r decrease, the marginal costs of production for both continuous importers and import starters from region r fall. Consequently, the price index P falls and so does the demand shifter B in (9). Due to this increased intensity of product market competition, not all the …rms that import some inputs from region r bene…t from the reduction in foreign input costs. As shown in (8) and (9), their operating pro…ts increase if and only if the increase in is is large enough that P 1 S s=1 s( 1) s is rises despite a fall in P . Import starters restructure their production networks. In particular, o¤shoring directly induces them to replace some domestic sellers (including themselves as input producers) with foreign sellers (an e¤ect that we refer to as the direct displacement e¤ ect). From Proposition 1, we learn that these newly added foreign sellers tend to produce the less di¤erentiated inputs. Thus, the displaced domestic sellers tend to be from the less di¤erentiated industries. In addition, import starters and continuous importers may restructure their production networks as a consequence of the reduction in their marginal costs (an e¤ect that we refer to as the productivity e¤ ect). Their operating pro…ts unambiguously increase relative to those of non-importers, since a reduction in P a¤ects all …nal-good producers equally.

23
The productivity e¤ect in turn a¤ects a …rm's domestic production networks through two channels. First, it follows from (10) that import starters have a greater incentive to expand search regions, relative to non-importers. O¤shoring entails a rise in is so that each region r's share of input varieties, given by isr = is , drops. Consequently, for each input industry, some import starters restructure their domestic supplier networks by adding distantly-located and productive sellers, while dropping the less productive ones in all other source regions (an e¤ect that we refer to as the within-industry restructuring e¤ ect). 37 Second, it follows from (11) that import starters also have a greater incentive to begin outsourcing inputs in a new industry, particularly the differentiated one that was not being sourced previously due to high variable trade costs (an e¤ect that we refer to as the industry composition e¤ ect). The following proposition summarizes the testable predictions about the various e¤ects of a fall in foreign input costs on the structure of newly o¤shoring …rms'domestic production networks.

Proposition 2
1. Relative to non-importers, import starters drop in every source region sellers that are on average less productive than others in the same industries. This extent of dropping is more profound in the newly-o¤ shored industries, since the direct displacement e¤ ect that some domestic suppliers are displaced by foreign suppliers is always present. Since the newly o¤ shored industries tend to be less di¤ erentiated, the dropped sellers in those industries tend to be located farther and more productive than those in other industries.
2. Relative to non-importers, import starters add sellers that are on average more productive and located farther than other …rms within each previously-outsourced industry. In addition, some import starters add sellers in industries from which they previously did not outsource inputs. Since such newly-outsourced industries tend to be more di¤ erentiated and thus entail higher communication costs, the sellers added in the newly-outsourced industries tend to be located closer than those in the previously-outsourced industries.
O¤shoring leads to the restructuring of …rms'domestic production networks, thereby a¤ecting industry coagglomeration. The direct displacement e¤ect induces coagglomeration as the sellers that are directly replaced by foreign sellers tend to be located farther as they produce inputs that are less di¤erentiated than sellers in other industries. The two types of productivity e¤ects of o¤shoring on industry coagglomeration are mixed. On the one hand, the within-industry restructuring e¤ect implies that import starters replace the less productive sellers with the more productive ones, which are located farther than others within the same industries. On the other hand, the industry composition e¤ect induces them to begin outsourcing inputs in the relatively more di¤erentiated industries so that they add sellers located closer than those in other industries. The following proposition summarizes these possibilities.
Proposition 3 Although the e¤ ects of o¤ shoring on industry coagglomeration are mixed, o¤ shoring induces industry coagglomeration if the within-industry restructuring e¤ ect is small relative to the direct displacement and industry composition e¤ ects.
The proof of Proposition 3 is relegated to Appendix A. The basic idea is that if the …xed sourcing costs are large so that within-industry restructuring in the newly o¤shored industries is limited, the direct replacement and industry composition e¤ects tend to dominate. Whether o¤shoring leads to industry coagglomeration depends on the relative strength of the three e¤ects of o¤shoring, which we examine empirically in Section 5.3.

Regression Analyses and Results
In this section, we empirically examine using the network data the three testable hypotheses derived in Section 4. For notational clarity, let us denote buyer, seller, industry (3-digit), and region (one of 47 prefectures) by i, j, s, and r, respectively. When industry and region …xed e¤ects are included in the regressions, we will be clear about whether they are for the buyer or seller.

Domestic Sourcing Patterns
We …rst examine Proposition 1, which is about the patterns of …rms'domestic sourcing. Equation (13) shows that …rm i's spatial pattern of domestic sourcing in industry s can be described by With these parameterizations, we can express the empirical counterpart of (14) as where an industry is de…ned as a JSIC 3-digit category. 38 F E srs(i) and F E sr stand for inputindustry reference-region and input-industry source-region …xed e¤ects, respectively. 39 With these …xed e¤ects included, we study the relationship between a buyer's scope of domestic sourcing in a region and the proximity of the source region (relative to the reference region from which it sources the same type of inputs). 40 To estimate and , we need estimates of the model's key parameters: s , s and s . We if a …rm's core productivity is distributed Fréchet with parameters T and , its revenue is also distributed Fréchet, with the location parameter equal to T A 1 and the shape parameter equal to =( 1), where A is a sector-speci…c variable. Therefore, we can use the mean and the standard deviation of …rm revenue in the TSR data to back out s (at the 3-digit JSIC level), given s . For s , we use the share of air freight costs in U.S. imports in industry s to proxy for the importance of timely delivery (see the Appendix for details). The idea is that if the delivery of a good is time sensitive, the slope of the variable trade cost with respect to distance is steeper. Table 4 reports the estimates of and , according to (15). Standard errors are clustered at the input-industry-source-region level. In columns 1 to 6, we use a buyer's nearest source region for each industry as the reference region, while in columns 7 through 9, we use a buyer's home region as the reference point. The cost of using a buyer's home region as the reference point is that not all buyers procure inputs in each sector from its home region. Thus, the number of observations will be smaller in the last three columns. When using the estimated s from Caliendo and Parro (2014), we …nd in columns 1 to 3 that a buyer's scope of domestic sourcing is decreasing in distance, more so for di¤erentiated inputs. Speci…cally, as reported in column 1, a 10% increase in the distance relative to the nearest region is associated with a 0.5% drop in the number of sellers for an industry that has the mean value of s (=9.82). 41 Column 2 shows that such negative correlation is more In columns 4 to 6, when we use the …rm-based estimate of s , we …nd quantitatively larger e¤ects.
According to the coe¢ cient on s s s 1 log (d ir ) in column 6, sectors that have one standard-deviation larger s = ( s 1) compared to the sectoral mean are associated with an additional 0.19% drop in the number of sellers from a region that is 10% away in distance relative to the reference region, evaluated at the mean value of s (=13.72). 43 The results become quantitatively more signi…cant when we use the buyer's home region as the reference region (see columns 7 to 9), or when we use the 2010 sample (see Table A5 in the appendix), or when we include the parent-children relationships (results available upon request).
In Table 5, we empirically examine the pattern of …rms' domestic sourcing at the extensive margin, by replacing the dependent variable in speci…cation (15) with a dummy for whether buyer i sources intermediates in industry s from source region r or not. Since we only have information on …rms'foreign sourcing at the broad sector level (12 manufacturing sectors) from BSJBSA, we consider a …rm's participation in domestic and foreign sourcing respectively at the broad sector level.
Information on …rms' domestic sourcing at the more disaggregated industry level is aggregated to the broad sector level. 44 Accordingly, we compute the weighted average of input industries' characteristics across 3-digit industries within a broad sector. The procedures of aggregation are described in detail in the data appendix. Since we no longer use information of the reference source region to de…ne the dependent variable, we drop F E srs(i) as a regressor, while retaining inputindustry source-region (F E sr ) …xed e¤ects. We sometimes include buyer …xed e¤ects as well to control for any unobservable determinants of domestic sourcing (e.g., buyer's productivity).
We use the estimated s from Caliendo and Parro (2014) in Panel A, and our own estimates using …rm-level data in Panel B. Regardless of the estimates, we …nd that a buyer's likelihood of sourcing inputs from a prefecture is decreasing in the bilateral distance (columns 1 and 7), after controlling for input-industry source-region …xed e¤ects. The negative correlation is stronger for the more di¤erentiated inputs (columns 2-3 and columns 8-9, respectively). These regression results remain robust and quantitatively similar after we control for buyer …xed e¤ects (columns 4-6 and columns 10-12, respectively). We …nd that the di¤erential e¤ect of distance on the incidence of domestic sourcing across industries is economically signi…cant. Based on the coe¢ cients reported in column 4, a 10% increase in distance relative to the home region is associated with a 0.1 percentagepoint decline in the likelihood of sourcing from the region, evaluated at the mean value of s . 45 Based on the estimates in column 5 and the mean value of s , the same distance is associated with an additional 0.1 percentage-point decline in the likelihood of sourcing for industries with a one standard-deviation larger s = ( s 1) relative to the sectoral mean. 46 In Table 6, we study the determinants of a …rm's decision to o¤shore inputs in an industry.
Without detailed information about the source country of o¤shoring, we examine Proposition 1 on o¤shoring by including a buyer's (log) TFP and the interaction term between (log) TFP and s = ( s 1). Our model predicts that more productive …rms are more likely to incur the …xed costs to o¤shore intermediates. Given that trade costs will be increasing in input industry's product di¤erentiation, such positive relationship should be weaker for the more di¤erentiated inputs. Re- 4 4 For instance, the dummy for a …rm's domestic sourcing is set equal to 1 for a broad sector if the …rm outsources in any 3-digit industry that belongs to the sector. 4 5 Using the coe¢ cient on ln(dist+1) s (= 0:001) and the mean value of s (=9:82) we come up with 0:098% = 0:001 10 9:82%. 4 6 Given that the mean and the standard deviation of s= ( s 1) are 1.328 and 0.262 (across 3-digit industries, see Table A9 in the  gardless of which estimated s used, we …nd a positive and signi…cant correlation between a buyer's likelihood of o¤shoring inputs and its productivity, after controlling for input-industry source-region …xed e¤ects. Such correlation is weaker for the more di¤erentiated input industries (see columns 1 and 3). These results remain robust even after we control for buyer …xed e¤ects in columns 2 and 4. We also …nd that a buyer's domestic sourcing in an input industry is positively related to the likelihood of o¤shoring in the same industry. This result implies that …xed costs to o¤shore could be lower if a …rm already incurs some of them for domestic sourcing. This pattern of sequential sourcing will be the basis for the construction of our instrument.

Relationship between Firms'Global Sourcing and Domestic Supplier Choices
The …nal part of the paper studies Propositions 2 and 3, which are about the e¤ect of …rms' o¤shoring on their choices of domestic suppliers. We …rst examine whether the likelihood of a buyer's dropping and adding domestic suppliers is associated with an exogenously induced o¤shoring decision. To this end, we estimate the following speci…cation using our two-year panel data on Japan's production networks (2005 and 2010): where I ij is a dummy variable indicating whether buyer i drops (or adds) seller j between 2005 and 2010. Each unit of observation is a buyer-seller link. The regression sample excludes all existing importers in 2005, as our goal is to study the e¤ect of the participation in o¤shoring, rather than that of the extent of o¤shoring.
When we run the "drop" regressions, we further restrict the sample by excluding each buyer's newly added seller links, such that we can gauge the e¤ect of o¤shoring on a …rm's decisions of dropping an existing domestic seller. In particular, the dependent variable of the "drop"regressions is de…ned as I ij = Drop ij , which is equal to 1 if seller j and buyer i were linked in 2005 but not anymore in 2010, and 0 otherwise (if the relationship continued).
When we run the "add"regressions, on the other hand, we restrict the sample by excluding each buyer's dropped seller links, such that we can empirically compare the characteristics between the buyer's continuing suppliers with those of its newly added suppliers, after it o¤shores inputs. The dependent variable of the "add" regressions is de…ned as I ij = Add ij , which is equal to 1 if a new link between buyer i and seller j is observed in the 2010 sample but not in 2005, and 0 otherwise (indicating an old relationship). 47 The variable imp i is a dummy indicating …rm i's switching from no o¤shoring in 2005 to o¤shoring any inputs between 2005 and 2010. 48 The regressor of interest is the interaction term x ij x i represents a measure of a supplier's characteristic, relative to the average characteristics of the sellers used by buyer i in 2005. We use a demeaned measure since a buyer compares the candidate supplier with its existing suppliers when making decisions to add or drop suppliers. For instance, a supplier considered to be large by a …rm may not be considered large by another …rm. More speci…cally, in the "drop" regressions, x i is constructed using the sample of sellers from which the buyer procured inputs in 2005. On the other hand, in the "add"regressions, x i is constructed using sellers from which the buyer procured inputs in 2010 (i.e., those that were dropped since 2005 will not be included in the construction of the buyer's mean). We consider two seller characteristics-log size (measured by either sales or employment) and log distance from a buyer. Buyer-industry (F E s(j) ), buyer-region (F E r(i) ), input-industry (F E s(j) ), and source-region (F E r(j) ) …xed e¤ects are always included to control for any region-and industry-speci…c trends of supplier adding and dropping, such as external economies of scale for di¤erent industries, initial levels of economic development, or local government policies. 49 To establish causality, we estimate speci…cation (16) where exp s;t is the global exports to all destination countries in the world except Japan. 50 To construct the instruments for o¤shoring at the …rm level, we exploit the information about a …rm's sectoral pattern of domestic sourcing. In Table 5, we …nd that …rms are more likely to o¤shore inputs in an industry in which it has already sourced inputs domestically. These results are suggestive of a lower …xed cost of o¤shoring, conditional on the …rm's domestic sourcing in the same industry. Thus, foreign countries'export supply shocks in an industry are more likely to induce a Japanese …rm to start o¤shoring if it has already sourced inputs in the same industry.
Based on these …ndings, we merge a …rm's sectoral pattern of domestic outsourcing with that of the vector of estimated export supply shocks, and compute …rm i's exposure to the world export supply shocks as where is is a dummy, which equals 1 if buyer i outsources industry-s inputs domestically in 2005, and 0 otherwise. We construct such dummies at the Japanese 4-digit SIC level (i.e., S = 563 manufacturing industries).
We use shock i speci…ed in (18) as an instrument for imp i , and shock i log x ij x i as an instrument for imp i log x ij x i , as suggested by Wooldridge (2010). Table 7 reports both the OLS and 2SLS results of estimating (16) for dropping, based on the sample of all buyers in the manufacturing sector that sourced inputs domestically but not from foreign suppliers in 2005. As reported in columns 1 to 4, while the OLS estimated coe¢ cient on imp i and imp i log x ij x i take the expected signs, they are all insigni…cant. 51 The 2SLS estimates in columns 5 to 8, however, show statistically signi…cant e¤ects of o¤shoring on …rms'dropping of domestic suppliers. In column 5, the estimated coe¢ cient on imp i shows that a buyer's o¤shoring decision reduces the likelihood of dropping domestic suppliers. Within buyer industries, buyer home regions, input industries, and source regions, a buyer that started importing since 2005 is on average 65% less likely to drop its existing sellers between 2005 and 2010. In column 6, we …nd that even though newly o¤shoring …rms tend to be less likely to drop domestic suppliers, they are relatively more likely to drop the larger one (columns 7 and 8). In particular, the coe¢ cient of 0.087 on the interaction term in column 7 implies that an existing supplier that is 50% larger than the mean supplier of a buyer is 4.3% relatively 5 1 The main endogeneity that we aim to tackle is a …rm's positive supply shock that may trigger its o¤shoring on the one hand, but a¤ect the probability of dropping domestic suppliers in either direction on the other. The insigni…cant results based via OLS do not contradict our signi…cant 2SLS results. 32 more likely to be dropped on average. If we measure …rm size by employment, the di¤erential e¤ect would be 6.5% according to the coe¢ cient of 0.129 on the interaction term reported in column 8.
All F statistics for the …rst stages of these 2SLS estimations suggest that our instruments pass the weak instrument test by a wide margin. Table A6 in the Appendix reports the regression results of the …rst stage. The coe¢ cients on the instrument and its interaction with seller characteristics are statistically signi…cant when the corresponding endogenous variables are used as the dependent variables of the …rst stage. Table 8 reports the estimates of the "add" regressions based on speci…cation (16). Columns 1 to 4 report the OLS estimates. It is not surprising that there is a positive correlation between a …rm's import dummy and likelihood of having a new domestic supplier. For instance, a positive productivity shock will induce a …rm to add both domestic and foreign suppliers simultaneously.
As reported in columns 5 to 8, the 2SLS estimates show that the newly o¤shoring …rms are less likely to add the relatively more distant domestic suppliers (column 6), but more likely to add the larger ones (columns 7 and 8).
The …ndings that newly o¤shoring …rms are more likely to add larger suppliers seem to contradict the earlier …ndings that they are also more likely to drop larger suppliers, as reported in Table 7.
However, through the lens of our model, these results can be rationalized by the joint force of the direct displacement, within-industry restructuring and industry composition e¤ects of o¤shoring.
The …ndings that larger domestic suppliers are more likely to be dropped are consistent with the direct displacement e¤ect. O¤shoring …rms' substitute domestic generic-input suppliers, which tend to be larger and more distantly located, with foreign suppliers producing similar inputs. The results that larger domestic suppliers are more likely to be added can be due to the within-industry restructuring e¤ect of o¤shoring that induces …rms to add more productive and distant domestic suppliers within each industry. The fact that more distant suppliers are less likely to be added can be explained by the industry composition e¤ect-di¤erentiated input industries, from which a buyer previously did not source inputs, are now being added to the buyer's sourcing set. Considering the rise in communication costs, the buyer will source inputs from closer domestic suppliers in the newly added, di¤erentiated industries. Table A7 in the Appendix shows that the coe¢ cients on the instrument and its interaction with seller characteristics are statistically signi…cant in the corresponding …rst-stage regressions.
To further study the three e¤ects of o¤shoring, in particular the industry composition e¤ect, we empirically examine the relationship between …rms' o¤shoring and the pattern of adding and dropping of input industries. To this end, we estimate the following speci…cation: where I is is a dummy variable indicating whether buyer i drops ( The variable of interest, imp i X s , is an interaction term between the change in buyer i's import status and industry s's product di¤erentiation (X s ), measured by either s = ( s 1) or the Rauch indicator, which equals 1 for di¤erentiated goods, and 0 otherwise (see Section B.1 in the Appendix for details). Buyer industry (F E s(i) ), home region (F E r(i) ), and input industry …xed e¤ects (F E s(j) ) are always included. Table 9 presents the estimates of (19) via OLS and 2SLS. By using shock i as an instrument for imp i , and shock i X s as an instrument for imp i X s , we …nd in our 2SLS regressions that newly o¤shoring …rms are less likely to drop (as measured by the Rauch indicator in column 5) but more likely to add the more di¤erentiated industries for domestic sourcing (both columns 7 and 8). These patterns of input industry adding and dropping support the model predictions, and are consistent with the hypothesis that the industry composition e¤ect and the direct displacement e¤ect are the main reasons for why closer suppliers are more likely to be added, while larger suppliers are more likely to be dropped, after a …rm o¤shores inputs. Table A8 in the Appendix reports the regression results of the …rst stage.

Concluding Remarks
In this paper, we study from both theoretical and empirical perspectives the spatial and sectoral patterns of …rms'global sourcing, as well as the e¤ect of o¤shoring on …rms'domestic production networks. We develop a multi-region global sourcing model in which …rms source inputs from suppliers in various input industries that di¤er in the degree of product di¤erentiation. Firms choose the optimal level of communication with suppliers, depending on the inputs'product di¤erentiation.
Using exhaustive data on buyer-seller links in Japan, we …nd that the more productive …rms source inputs from more suppliers and domestic regions, including the more distant ones. Distant suppliers are more productive on average, while productive …rms are more likely to o¤shore inputs. The negative correlation between distance and the extent of domestic sourcing is stronger for di¤erentiated inputs. Using a …rm-level instrument based on the sectoral patterns of buyers' initial domestic sourcing and global export supply shocks, we study the causal impact of a …rm's o¤shoring on its choices of domestic suppliers. We …nd that upon o¤shoring, …rms are less likely to drop domestic suppliers on average, but are relatively more likely to drop the larger ones. Newly o¤shoring …rms tend to add domestic suppliers that are relatively larger but also those that are more proximate. These …ndings suggest that the within-industry restructuring and industry composition e¤ects of o¤shoring, both arising from declines in marginal costs, play important roles in shaping …rms'domestic production networks. The resulting reorganization of …rms'domestic supplier relationships reduces the average distance between buyers and sellers, increasing the spatial concentration of production clusters in Japan.
Lower communication and transportation costs obviously have made production more global.
Our paper shows that while input sourcing has become more spatially dispersed across countries, it may lead to localization of production within a nation or region. These results echo the World Given a distribution function of the prices for the input varieties for every sector s, which is denoted by G isr (p), …rm i chooses the input levels fx isr (p)g p2[0;1) to minimize the cost of producing composite input s. The unit cost function for …rm i is the solution to the minimization problem:

1:
By solving this problem, we obtain the optimal input levels as ; for r 2 is : Substituting these solutions to P r2f0g[ is (I isr ) R 1 0 px isr (p)dG isr (p) gives us the unit costc is given in (3).

Sets of Source Regions
The unit cost of inputs equals p = z 1 w 0 c s for insourced inputs, while it equals where isr is de…ned in (5).
The input-price (or input-cost) probability distribution for every input variety of type s is common across source regions and can be written as and hence the unit cost of the composite input s is given byc is = s

A.3 Sourcing Strategy and Industry Equilibrium
As AFT points out, buyers' choice of source regions requires some consideration as to whether the choice of individual source regions exhibits substitutability. To see this, we examine how an addition of a source region, say region r 2 , a¤ects the sourcing potential of another source region, say region r 1 , within the same input industry by taking a further di¤erence of the expression in (10): We see immediately that the pro…t function is supermodular in isr 1 and isr 2 if s ( 1) > s , which AFT call the complements case, while it is submodular if s ( 1) < s , which is called the substitutes case. Adding a source region increases the sourcing potential of other regions in the complements case, while it decreases the sourcing potential of other regions in the substitutes case.
In contrast, the …rst-order approximation of the impact of an inclusion of region r 2 as a source 38 region in sector s 2 on the sourcing potential of region r 1 in another input sector, say s 1 , is given by In this case, the pro…t function is unambiguously supermodular. Adding a source region always increases the sourcing potential of any region for all other input sectors.
As argued by AFT, buyers'choice of source regions is rather simple if s ( 1) > s for all s, since adding a region will never give the buyers incentive to drop any existing source regions. In other cases where s ( 1) < s for some s, however, it is possible that r 1 6 2 is while r 2 2 is even though isr 1 > isr 2 ; this can arise if an inclusion of r 1 would lead to an exclusion of some other regions from is while an inclusion of r 2 with a small isr 2 will not.
There is a unique industry equilibrium in this model. For a given B in (9), each buyer i optimally chooses the set of source regions f is g S s=1 . The unit cost of …nal good production is determined accordingly as shown in (7), which in turn determines the price index P and hence an associated value of B, say B 0 , as shown in (9). Let

A.4 Proof of the Propositions
Potential source regions di¤er from one another in various aspects such as the number of inputproducers, technological level, and wage rate. To isolate the distance e¤ect, we assume here that all parameters other than s take the same values across di¤erent input industries. Omitting the subscript s for those parameters and also omitting the …nal-good producer index, for notational simplicity, the sourcing potential given in (5) can be written as Furthermore, equations (12) and (13) can be rewritten as s ; for any r; (22) We also assume here that source regions are complements, i.e., ( 1) > , to conduct a rigorous analysis. As argued in the previous subsection, we are not able to perfectly predict equilibrium sets of source regions in the substitutes case where ( 1) < . In the following analysis, we focus on the complements case and use a nice property that when sr 1 > sr 2 , if sr 2 2 s , so is sr 1 .

A.4.1. Proposition 1
We see from (21) that sr falls with d r for r 6 = 0. Moreover, the elasticity of sr with respect to m(d r ) and t(d r ) are s =( s 1) and t(d r ), respectively, which indicates that the adverse distance e¤ect is greater for the more di¤erentiated inputs.
To see how the distance e¤ect on the ranking of potential source regions varies with the degree of input di¤erentiation, we …nd from (23) that the distance plays a bigger role for the more di¤erentiated inputs, i.e., the inputs associated with a greater s =( s 1), in ranking the regions. At the one extreme where s =( s 1) ! 1, (23) shows that distance is just one of the factors that a¤ect relative attractiveness of a region. At the other extreme where s =( s 1) ! 1, we see that distance is the only factor that a¤ects the ranking of the source regions. So we infer that close regions are more likely to be ranked higher than farther regions. Letting k s (r) denote region r's ranking as a source region in industry s, we indeed have that if k s 1 (r 1 ) < k s 1 (r 2 ) and k s 2 (r 1 ) > k s 2 (r 2 ) for some r 1 and r 2 and for some s 1 and s 2 such that s 1 < s 2 (i.e., inputs in industry s 1 are less di¤erentiated than those in industry s 2 ), then we have d r 1 > d r 2 . Close regions are more likely to ranked higher so that more likely to be chosen as source regions in industries for the more di¤erentiated inputs.
The share of region r 1 relative to insourcing and that of region r 1 relative to region r 2 are given by (22) and (23), respectively. To see how they vary with the degree of inputs di¤erentiation, we take a logarithm and di¤erentiate them with respect to s =( s 1): where the …rst inequality obtains since s =[( s 1)m(d r 1 )] = q sr 1 is a probability and hence is less than 1. These suggest that the share of insourcing relative to any source region and the share of a region relative to another, farther region are both higher for the more di¤erentiated inputs.
Indeed, we can further show that the share of insourcing itself is higher for the more di¤erentiated inputs. To this end, we de…ne r s j and k s as the j-th region in the ranking and the optimal number of source regions for industry s, respectively, and rewrite (10) and (11) as Figure A2 shows how the optimal sets of source regions are determined for industries s 1 and s 2 , where s 1 < s 2 . It follows from (26) that the intersections, illustrated as s 1 and s 2 in the …gure, give us the optimal sourcing decision of the …rm for the two input industries, ignoring the integer problem (which can be justi…ed especially when the number of source regions is large so that smallest sourcing potential sr in s is small). If both upward-sloping and downward-sloping curves for industry s 1 are located above those for industry s 2 , as illustrated in the …gure, the optimal sourcing capability for s 1 is greater than that for s 2 , i.e., s 1 = s 1 0 + P ks 1 j=1 s 1 r s 1 j > s 2 0 + P ks 2 j=1 s 2 r s 1 j = s 2 , since s 1 0 = s 2 0 as indicated in (21). It is easy to see that this will be the case if s 1 r s 1 j > s 2 r s 2 j for any j.
The inequality s 1 r s 1 j > s 2 r s 2 j can be shown from the observation that s 1 r > s 2 r for any r, which in turn follows from (24). To this end, we consider a series of rankings for industry s 2 , where any consecutive rankings are di¤erent in a permutation of two regions, such that the region with a larger sourcing potential moves up in ranking while the one with a smaller sourcing potential moves down. The series starts with the ranking for industry s 1 and ends with the ranking for industry , followed by fr j(1) g M +M j(1)=1 , and so forth, and end with fr j(n(s 1 ;s 2 )) g M +M j(n(s 1 ;s 2 ))=1 = fr s 2 j g M +M

j=1
, where n(s 1 ; s 2 ) denotes the number of permutations necessary to reach from fr s 1 j g M +M j=1 to fr s 2 j g M +M

j=1
. We shall show s 1 r s 1 for any h 2 f0; 1; ; n(s 1 ; s 2 )g. We begin with the observation, from (24), that s 1 r j(0) = s 1 0 > s 2 r j(0) = s 2 0 . Suppose then that s 1 r s 1 j = s 1 0 > s 2 r j(h) = s 2 0 for any h, and show that the counterpart inequality also holds for h + 1. In the (h + 1)-th step, the permutation of region r, which used to be in the l-th place, and region r 0 , which used to be in the l 0 -th place, occurs such that where the …rst inequality follows from l < l 0 while the second inequality follows from the supposition for the h-th step. As for the l 0 -th place, we have Having established s 1 > s 2 , it follows immediately from s 1 0 = s 2 0 that s 1 0 = s 1 < s 2 0 = s 2 , i.e., the share of insourcing is greater for the industry with higher inputs di¤erentiation.
This ends the proof of Proposition 1.

A.4.2 Proposition 2
We shall show that the less di¤erentiated inputs are more likely to be outsourced and then show the direct and indirect impacts of o¤shoring on dropping and adding of input sellers.
We show here that for s 1 < s 2 , if the …rm outsources in industry s 2 , it also outsources in industry s 1 . The converse is not true. To show this claim, we use which has been established in the proof of Proposition 1. If k s 1 k s 2 , on the one hand, then we where the …rst inequality follows from (28) while the second inequality follows from k s 1 k s 2 . If k s 1 > k s 2 , on the other hand, we have Now, consider the case where the …rm begins importing some of the inputs in some industry, say s 1 , from a foreign region, say r . The direct consequence of this is a rise in s 1 . As a result, the share of input varieties sourced from every other source region, s 1 r = s 1 , drops. The dropped input suppliers tend to be less e¢ cient.
As s 1 increases due to o¤shoring, the unit cost of the …nal good decreases for all import starters, as shown in (7). As (8) indicates, this gives all the import starters incentive to expand the set of source regions and also to expand the industries in which the …rms outsource some of the inputs. Although the demand shifter B is negatively a¤ected, as shown in (9), the import starters have incentive to expand the search for input suppliers relative to the non-importers, since a decline in B a¤ects all …nal-good producers equally. Therefore, the import starters add some regions to the set of source regions for some industries, while they drop, as a consequence, ine¢ cient suppliers from all the existing source regions in those industries. Since newly-added regions are associated with the smallest sourcing potentials compared with the existing source regions within the industries, the newly-added sellers tend to be distantly located and more productive. If an import starter begins outsourcing some of the inputs in an industry, then we know from the above claim that the industry produces the most di¤erentiated inputs of all the industries in which the …rm outsources. Proposition 1 tells us that the sellers in that industry tend to be closer than sellers in other industries. This ends the proof of Proposition 2.

A.4.3 Proposition 3
Proposition 3 and the discussion that follows claim that if the …xed costs of outsourcing and those of searching regions for input suppliers are large, the o¤shoring tend to result in industry coagglomeration.
Proposition 2 shows that …rms tend to o¤shore less di¤erentiated inputs. Those inputs tend to be outsourced from distant regions, so import starters tend to replace distant suppliers with foreign sellers. Thus, the direct displacement induces industry coagglomeration. Proposition 2 also shows two indirect channels through which the average distance between buyers and sellers is a¤ected.
The …rst channel is the within-industry restructuring e¤ect. Import starters tend to add suppliers in distant regions while dropping …rms in every existing source regions, so that the reshu-ing induces dispersion within the industries. The other channel, however, is for industry coagglomeration. It is the channel of adding new ones in the list of the outsourcing industries, which we have referred to as the industry composition e¤ect. As we have seen, such industries tend to produce the more di¤erentiated inputs, and hence the sellers, which are newly-added, are located closer than those in other industries on average.
If f and f s are large, the indirect e¤ect tends to be for industry coagglomeration, since then the number of outsourcing industries is small so that the dispersion e¤ect is relatively small. In addition, the indirect e¤ect itself becomes small relative to the direct e¤ect, since the import starters tend to expand the search for new input suppliers less aggressively in that case, as we can see from (8). Thus

Product Di¤erentiation
Source: Rauch (1999). Description: a dummy variable that takes value equal to 1 for di¤erentiated products, and 0 for homogeneous products. There are two versions of the di¤erentiation 45 indicator-"conservative" and "liberal". In the main regressions, we use the "conservative" version, which has a lower number of commodities classi…ed as either organized exchange or reference priced. Since industries are originally de…ned at the 4-digit level based on the ISIC (revision 3) classi…cation, we use the following procedures to construct the di¤erentiation dummies for each JSIC 3-digit category. 53 1. We …rst map each ISIC 4-digit code (292 categories) to multiple JSIC (rev. 11) 4-digit codes The unique match is determined based on the number of HS 10-digit categories shared between a pair of SITC and HS code. For a handful of cases that a unique match cannot be determined because there are multiple pairs that tie in terms of the number of HS 10-digit categories shared, we will pick the smaller SITC in terms of the numeral value, albeit arbitrarily. Since a SITC may be split into multiple HS 4-digit category, we will split the export value of a SITC category using the number of HS 10-digit categories shared with each HS 4-digit as weight. We then repeat the same procedure to match export values at the HS 4-digit levels to each IO code (361 of them), using the concordance …le from the Statistics Bureau of Japan, and …nally map export values from IO categories to JSIC (rev. 11) 4-digit categories, using another concordance …le from the Statistics Bureau. We then construct a …rm-level instrument as speci…ed in Equations (17) and (18) in the main text.

B.3 New Method to Estimate the Dispersion of Firm Productivity ( )
Firm i's revenue can be written as R(z i ) = Az 1 i , where A is a constant that is common across all …rms in the industry. 54 The distribution of a …rm's revenue, R, can be written as where F denotes the cumulative distribution function of Fréchet(T , ), implying that R is also distributed Fréchet with the location parameter equal to T 0 = T A 1 and the shape parameter

1:
Using Mathematica, we can identify 0 using 2 and 2 for each 3-digit JSIC industry, each computed using …rm revenue data from TSR Company Information Database. Given estimates of s from Soderbery (2015), we can then compute s as ( s 1) 0 s , for each industry s. 5      Note: The regression sample includes manufacturing buyers only and domestic suppliers from both manufacturing and non-manufacturing industries. Each observation is a buyer. When constructing the buyer-specific measures of domestic sourcing, parent-child relationships and sellers with fewer than 5 employees are dropped. The number of observations in column 7 is significantly smaller because not all buyers added or dropped sellers during the sample period. A buyer's TFP is estimated using the Olley-Pakes method with the buyer's value added as the dependent variable. Robust standard errors, clustered by the buyer's region, are used. All existing importers in 2005 are excluded in the sample, so only import starters and non-importers are included. ***, **, * indicate significance at the 1%, 5%, and 10% levels, respectively.  Table A5 in the appendix. Observations are at the buyer input-industry source-region level. All regressions include input-industry-closest-region and input-industry-source-region fixed effects, where the closest region is the closest prefecture from which firm i sources intermediate inputs in a particular industry. Standard errors, clustered at the input-industry-source-region level, are reported in parentheses. ***, **, * indicate significance at the 1%, 5%, and 10% levels, respectively.  (2014), while those in Panel B use our own firm-based estimates of θ. See Section B.5 in the appendix for the detailed estimation procedures. All columns include input-sector-sourceregion fixed effects, while columns 4-6 and 10-12 include buyer fixed effects as well. Standard errors, clustered at the buyer level, are reported in parentheses. ***, **, * indicate significance at the 1%, 5%, and 10% levels, respectively.

Caliendo-Parro Estimates
Estimates using Firm Data  (2014), while columns 3 and 4 use our own firm-based estimates of θ. All columns include input-sector fixed effects, while columns 2 and 4 include buyer fixed effects as well. Standard errors, clustered at the buyer level, are reported in parentheses. ***, **, * indicate significance at the 1%, 5%, and 10% levels, respectively.

Offshore input industry Caliendo-Parro Estimates
Estimates using Firm Data The sample includes only manufacturing buyers that did not import in 2005. Newly added sellers are removed from the sample. The unit of observation is a buyer-seller pair. Parent-child relationships are removed from the sample. In column 5, the dependent variable of the first stage is a buyer's import starting dummy. For each regression in columns 6 to 8, there are two first stages, with the dependent variable being a buyer's import starter dummy or its interaction with a seller characteristic. The instrument for a buyer's import starter dummy is the variable shock i constructed based on Equation (18) in the text, while the instrument for the import starter interaction is shock i interacted with the corresponding seller characteristic. See Table A6 for the regression results of the first stage. Robust standard errors are reported in parentheses. ***, **, * indicate significance at the 1%, 5%, and 10% levels, respectively.

OLS 2SLS
The sample includes only manufacturing buyers that did not import in 2005. Dropped sellers are removed from the sample, so that the comparison is between new suppliers and continuing suppliers. The unit of observation is a buyer-seller pair. Parent-child relationships are removed from the sample. In column 5, the dependent variable of the first stage is a buyer's import starting dummy. For each regression in columns 6 to 8, there are two first stages, with the dependent variable being a buyer's import starter dummy or its interaction with a seller characteristic. The instrument for a buyer's import starter dummy is the variable shock i constructed based on Equation (18) in the text, while the instrument for the import starter interaction is shock i interacted with the corresponding seller characteristic. See Table A7 for the regression results of the first stage. Robust standard errors are reported in parentheses. ***, **, * indicate significance at the 1%, 5%, and 10% levels, respectively. The sample includes only manufacturing firms that did not import in 2005. The unit of observation is a buyer-input-industry pair. The dependent variable of the regressions in columns (1), (2), (5) and (6) is a dummy variable equal to 1 if a 3-digit industry was dropped by a buyer between 2005 and 2010, 0 otherwise. The dependent variable of the regressions in columns (3), (4), (7) and (8) is a dummy variable equal to 1 if a 3-digit industry was added by a buyer between 2005 and 2010, 0 otherwise. For each regression in columns 5 to 8, there are two first stages, with the dependent variable being a buyer's import starter dummy or its interaction with the input industry's characteristic. The instrument for a buyer's import starter dummy is the variable shock i constructed based on Equation (18) in the text, while the instrument for the import starter interaction is shock i interacted with the corresponding input industry's characteristic. See Table A8 for the regression results of the first stages. Robust standard errors are reported in parentheses. ***, **, * indicate significance at the 1%, 5%, and 10% levels, respectively.        Buyers' (4-digit) Industry Note: The regression sample includes manufacturing buyers only and domestic suppliers that are either manufacturing or non-manufacturing. The unit of observation is at the buyer level from columns (1) to (6), and at the buyerseller level in columns (7). All regressions include the most exhaustive set of fixed effects possible. Standard errors, clustered at the buyer's industry level, are reported in parentheses. ***, **, * indicate significance at the 1%, 5%, and 10% levels, respectively. Note: The regression sample includes manufacturing buyers only and domestic suppliers from both manufacturing and non-manufacturing sectors. Parent-child relationships are removed from the sample. Data for 2010 are used. The unit of observation in all columns is at the buyer-source-region-sector level. All regressions include input-industry-closest-region (or input-industry-homeregion in columns (4)-(6)) and input-industry-source-region fixed effects, where the closest region is the closest prefecture from which firm i sources intermediate inputs in a particular industry. Standard errors, clustered at the industry-source-region level, are reported in parentheses. ***, **, * indicate significance at the 1%, 5%, and 10% levels, respectively.

Bottom 20 Input Industries in terms of Average Distance from Buyers
Note: "reg" denotes the regression sample.