WHO PRODUCES "MADE IN CHINA": CHINA S INTERNAL MIGRATION AND EXPORT GROWTH

WHO PRODUCES "MADE IN CHINA": CHINA S INTERNAL MIGRATION AND EXPORT GROWTH Chen Liu Xiao Ma University of California, San Diego December 13, 2017 Abstract Rapid growth in productivity and the liberalization of trade barriers are viewed as the major determinants of growth in China s exports, which have increased by a factor of more than 20 in real terms since the early 1980s. This paper quantifies the contribution, previously unknown, of various sources of export growth. We analyze determinants that have commonly been associated with export growth but add a less traditional mechanism, that is, China s massive internal migration, which has lowered labor costs in exporting sectors and reinforced China s export advantages. Using China s 2005 population survey, we first document that China s internal migrants are a crucial labor force in China s exporting provinces and in manufacturing sectors. We then employ a spatial general equilibrium trade model to account for the proportion of China s export growth due to each source. We find that productivity growth accounts for 50-60% of export growth between 1990 and 2005. Tariff reduction explains about 26% of export growth, whereas internal migration resulting from reduced barriers to labor mobility accounts for another 15%. We also find that without the reform to internal migration policy, China s export activity would have been replaced disproportionately towards East Asian and Southeast Asian countries. We are grateful to participants at the UC San Diego graduate student lunch seminar for helpful comments. We would also like to thank Robert Feenstra and the Center for International Data at the University of California, Davis for sharing data with us on China s customs transactions for the period 1988-1991. 1

1 Introduction China s rise as the "global factory" has been one of the most significant economic phenomena in the last three decades. From the early 1980s to 2005, China s exports grew by a factor of 23 in real terms; during that period, the share of global expenditure on manufactured goods "Made in China" grew from 0.8% to 13%, and to 18% among rich economies. A large literature has emerged that studies the impacts of China s spectacular export growth on the United States (Autor, Dorn and Hanson, 2013) and from a global point of view (Di Giovanni et al., 2014; Hsieh and Ossa, 2016). Economists have traditionally considered the major source of China s export boom to be productivity growth due to factors such as the decentralization of state-owned enterprises (Hsieh and Song, 2015; Huang et al., 2017); rural-urban migration (Tombe et al., 2015); the entry of highly productive firms (Brandt et al., 2012); access to foreign technologies and capital (Hsieh and Klenow, 2009); and productivity growth associated with trade liberation (Khandelwal et al., 2013; Brandt et al., 2017). 1 The liberalization of trade barriers in the form of reducing trade uncertainty (Handley and Limão, 2017) or the elimination of export intermediaries (Bai, Krishna and Ma, 2017) has also directly boosted China s export growth. In addition, China s enormous number of internal migrants has been viewed as the backbone of the country s industrial labor force (Li, Li, Wu and Xiong, 2012). However, the quantitative importance of each source s contribution to export growth remains unknown. In this paper, we quantify the extent to which these factors promote China s export growth and explore the importance of China s massive internal migration resulting from a reduction in barriers (the Hukou reform). Historically, population mobility in China was strictly controlled through the household registration system (Hukou). As a part of China s transition to a market economy, regulations have been greatly relaxed since the 1990s. As a consequence, over 150 million Chinese workers have migrated from less developed interior provinces to the more industrialized coastal provinces (Chen et al., 2010). We begin by documenting three facts regarding China s internal migration and exports. First, China s domestic migrants, the majority of whom migrate from interior provinces to coastal provinces, have constituted the crucial manufacturing labor force in coastal provinces. In Guangdong Province, migrant workers account for 56% of the provincial manufacturing labor force. Second, provinces that face larger increases in their migrant labor force in manufacturing sectors experience faster growth rates in manufacturing exports. Third, we show that China s rapid increase in internal migration precedes China s export surge. This evidence 1 Khandelwal et al. (2013) examine the productivity gains associated with the removal of quotas on Chinese textile and clothing exports. Yu (2015) shows that reduced tariffs on both imported inputs and final goods affect the productivity of large Chinese trading firms. Brandt et al. (2017) estimate the imact of cutting tariffs on productivity. 2

suggests that by providing cheap labor, China s massive internal migration has played a substantial role in promoting China s manufactured export growth. We then employ a spatial general equilibrium trade model to quantitatively investigate the causes and consequences of China s export growth. The production side of our model extends from a multi-sector Melitz-Chaney model (Arkolakis et al., 2017), although we incorporate input-output linkage (Caliendo and Parro, 2015) and allow agglomeration forces. Firms in each country can choose where to open an establishment and whether to serve each destination market. The labor supply side is borrowed from the Roy-like model (Lagakos and Waugh, 2013; Hsieh et al., 2013), and we extend it to consider endogenous location and sector sorting of internal migration: Workers choose locations and sectors to work in by fully recognizing the rewards to skills in each market as well as the amenities and migration frictions. The model captures policy-related migration barriers as destination- and sector-specific migration costs. Reduction of migration barriers due to the Hukou reform incentivizes workers to migrate to coastal provinces for higher wages. This leads to an increase in labor supply in China s coastal provinces and a more disproportionate increase in the labor supply for manufacturing sectors. As a consequence, wages fall more in manufacturing sectors in coastal provinces. Following the insight from Melitz (2003), the decline in labor costs promotes China s export growth through two channels: the intensive margin of export, that is, exporting firms export more intensively as the lower price of China s export goods makes them more competitive globally; and the extensive margin of export, that is, more firms are established to export in coastal provinces in response to reduced labor costs in those provinces. In addition, the model also allows for agglomeration forces, that is, as more manufacturing firms cluster in coastal provinces, the sector s total productivity factor rises. We use the model, together with detailed data on China s internal migration and bilateral trade flows, to account for China s export growth since 1990 resulting from three sources: productivity growth relative to the global economy, tariff reduction, and decreased barriers to internal migration. 2 We quantify these impacts by performing three counterfactual exercises: The first one only differs from the data in that China s relative productivity growth is absent, the second one only differs from the data in that tariff reductions are absent, and the third only differs from the data in that migration barriers are not reduced. We attribute the unexplained component of export growth to the reduction in trade barriers that is not captured by tariff changes. Our counterfactual exercise requires measurement of changes in migration frictions in each sector and each destination province from 1990 to 2005. We estimate changes in destination- 2 The source of China s TFP growth considered in this accounting exercise is a combination of sources that have driven productivity growth; for example, the resource reallocation resulting from privatization of stateowned enterprises; access to foreign capital and technology; and productivity growth induced by a reduction in trade barriers. 3

sector migration frictions from changes in origin-destination-sector migration labor forces relative to the changes in the labor forces of those who stay in their home province and in the same sector, while netting out observable changes in transportation costs and economic conditions, which also affect migration decisions. 3 We also measure changes in China s TFP growth from 1990 to 2005. Combining China s customs transaction database and bilateral trade flow from the Comtrade database, we perform a gravity-type estimation to measure changes in China s TFP growth resulting from changes in provincial-sector exports to the rest of the world. Combining microcensus data for multiple countries and trade data from China s customs transaction database and UN Comtrade, we calibrate a model that includes 30 Chinese provinces, 34 foreign countries and the rest of world. 4 We find that China s Hukou reform explains 15% of China s export growth since 1990. Unsurprisingly, productivity growth is the major source, accounting for 50-60% of export growth. Tariff reduction accounts for about 26% of the country s export growth. We further investigate how internal migration resulting from liberalization of barriers to internal labor mobility explains differential export growth to different foreign countries and sectors. We find that internal migration is responsible for more of China s export growth to developed countries and contributes more to export growth in the computer, electronic & optical equipment sectors. 5 This finding is driven by the following two facts: First, developed countries import more from Chinese provinces in which manufacturing sectors employ migrant workers more intensively. Second, the computer, electronic & optical equipment sectors employ migrant workers more intensively. We find that if China s restrictions to internal labor mobility had not changed, its export activity would have been disproportionately reallocated to East Asian and other Southeast Asian countries. For instance, 24.6% of China s migrationinduced export growth would have been reallocated to East Asian countries such as Korea and Taiwan, and 11.0% of China s migration-induced export growth would have been reallocated to Southeast Asian countries. This paper is related to the literature on Chinese export growth and its associated welfare and employment consequences for the US (Autor et al., 2013). Within this literature, a substantial body of research has emphasized the global welfare implications of China s productivity growth (Di Giovanni et al., 2014; Caliendo, Dvorkin and Parro, 2015; Hsieh and Ossa, 2016), whereas other authors have emphasized the impacts of various forms of trade liberal- 3 We parameterize transportation costs as a function of geographic distance. The identification of changes in migration frictions requires assuming that the migration friction of staying in one s home province is unchanged over time. 4 We solve the model in changes by using the exact hat algebra, which requires only two parameter values: the migration elasticity and the trade elasticity. We obtained these two elasticities from previous literature. 5 We find that internal migration explains 18.7% of China s export growth to France, 17.7% to United Kingdom and 17.7% to the United States; it also explains 21.0% of China s export growth in the computer, electronic & optical equipment sectors, and 20.4% in the electrical machinery & apparatus sector. 4

ization on export growth (Brandt and Morrow, 2017; Bai et al., 2017) 6, and on productivity growth (Khandelwal et al., 2013; Yu, 2015; Brandt et al., 2017). This paper complements the literature by accounting for the extent to which each source has contributed to export growth. We also emphasize a quantitatively less well-understood factor behind China s growth in exports: namely, reduced barriers to labor mobility and the associated massive internal migration. What makes this case special is that China has historically regulated population mobility. By analyzing internal migration resulting from the Hukou reform, our analysis shows that in a highly globalized world economy, liberalizing institutional frictions in a large country can generate substantial global impacts through international trade. This paper also relates to the literature which jointly analyzes China s internal migration and trade. China s internal migrants have been studied as an important key to alleviation of the misallocation of labor in China and promotion of aggregate productivity growth (Tombe et al., 2015; Ma and Tang, 2016). Fan (2015) analyzes how domestic migration frictions affect the aggregate and distributional welfare impacts of trade liberalization. Zi (2016) focuses on a special form of trade policy China s reduction of the input tariff and argues that Hukou reform, by relaxing population mobility barriers, intensified China s welfare gains from greater openings to trade. Our analysis differs from previous studies in two aspects. First, previous studies analyze the spatial sorting of China s internal migrants, but abstract from the sectoral sorting. We consider both the spatial and sectoral sorting of China s internal migration. By showing that China s massive internal migration preceded its export surge, we document an important mechanism that is missing in the literature, that is, China s internal migration has provided a crucial manufacturing labor force and has promoted China s export growth. Second, our paper studies a different set of internal migrants: those who work in a province other than the place of their Hukou registration. 7 This allows us to measure the size of province- and sector-specific migration that is the consequence of Hukou reform during the period 1990-2005. 8 This paper is also related to the broad literature using quantitative trade models, for example, Eaton and Kortum (2002) and Melitz (2003), while incorporating elastic labor supply and geographic frictions (Allen and Arkolakis, 2014; Caliendo, Dvorkin and Parro, 2015; Redding, 2016; Monte et al., 2015) to analyze the welfare consequences of aggregate economic shocks. 6 Brandt and Morrow (2017) show that a decline in import tariffs leads to an increase in province-sector export volume. 7 Tombe et al. (2015) consider both inter-provincial migrants and rural-urban migrants during 2000-2005; they define rural-urban migrants as the sector mismatch between the Hukou and the actual employment. As the nonagricultural sector has developed rapidly in rural China (more than 50% of rural employment was in sectors other than agriculture in 2000), the rural-urban migration measured in Tombe et al. (2015) is subject to the confounding of urbanization. Fan (2015) examines pre-2000 internal migrants who are defined as the mismatch between workers the place of residence and the place of birth. 8 Zi (2016) also measures internal migration induced by Hukou reform based on a 5-year migration intensity. Our measurement differs in that it is based directly on self-reported data on province of Hukou registration and also provides information about migrants employment sector. 5

The production side of our model is built on Arkolakis et al. (2017), while also taking into account the input-output linkages across industries (Caliendo and Parro, 2015). The labor supply side of the model is built on the increasingly prevalent Roy-like assignment model, which has been developed to analyze agriculture and non-agriculture productivity differences (Lagakos and Waugh, 2013); labor-to-occupation misallocation and its implications for aggregate productivity (Hsieh et al., 2013); computerization s effects on inequality (Burstein et al., 2015); and trade-induced wage inequality (Galle et al., 2015; Adão, 2015; Lee, 2016). We extend this model to incorporate frictions in labor mobility. The paper is organized as follows: Section 2 discusses the background of China s Hukou reform, and presents facts which relate China s internal migration with international trade; Section 3 represents the quantitative model. Section 4 discusses the data source used in this paper, and presents the estimation of migration frictions and productivity growth. Section 5 discusses the quantitative exercise performed in this paper; Section 6 analyzes the role of China s internal migration. Section 7 concludes. 2 Background Launched in the late 1950s, China s Residence Registration System (Hukou) was designed to control internal migration flows. The system assigns each household a Hukou, based on the residential location. A Hukou defines the geographic area in which a Chinese citizen is eligible to work and receive public benefits. The movements of internal migrants, including those who migrate without a local Hukou (non-hukou migrants) and those who acquire a local Hukou after migrating (Hukou migrants), were tightly controlled in the years following the system s implementation. The barriers to internal labor mobility gradually began to be relaxed in the early 1980s. An important change was that the government allowed people to live and work in locations other than their home city by obtaining a temporary residence permit. A series of deep reforms took place in the late 1990s. Until 2003, the main aspects of changes in the Hukou were: decentralization of the Hukou enforcement power to local government, abolition of the ruralto-urban migration quota, elimination of the temporary residence certificate requirement, and an easing of Hukou conversion for elderly parents, highly educated workers, skilled workers and investors. These policy changes have triggered the movement of over 150 million internal migrants. One challenge of measuring China s internal migration is related to the confounding of China s rapid pace of urbanization, as rural-urban geographic boundaries have quickly shifted over the last two decades. As a result, changes in workers residence type (urban or rural) may not be due to spatial movement per se, but may also derive from urbanization. This paper 6

uses geographically consistent boundaries to measure the set of internal migrants of interest. We focus on inter-provincial non-hukou migrants, whose Hukou registration province differs from the province where they currently work. Non-Hukou migrants are often referred to as the "floating population" in China, and their massive emergence is one of the major consequences of Hukou reform. Hereafter in this paper, we refer to inter-provincial non-hukou migrants as internal migrants. In this section, we use China s 2005 Population Survey 0.2% sample for data on China s internal migration. We restrict the sample to individuals who are between 20 and 60 years old and not attending school. Section 2.1 documents the observed characteristics and the spatial movement of non-hukou migrants of interest. Section 2.2 documents a less well-known situation: that is, that non-hukou migrants constitute a crucial manufacturing labor force in coastal provinces. Section 2.3 establishes a positive correlation: Provinces that have faced large increases in the migrant labor force in their manufacturing industry since 1990 have experienced faster growth in manufactured exports. Section 2.4 shows that the rapid migration growth occurred before growth in exports. 2.1 The Spatial Allocation and Characteristics of Internal Migrants We document the spatial movement of China s internal migrants in Table 1, which shows the percentage of the total number of internal migrants represented by each group of non-hukou migrants who moved between a specific origin and destination. Looking at column (1), as might be expected more than 78% of all inter-provincial non-hukou migrants are bound either for the highly developed coastal provinces or Beijing. Guangdong Province alone absorbs 35.6% of the total migrants. The share received by other coastal provinces is also substantial: 13.0% for Zhejiang, 9.4% for Shanghai, 8.7% for Jiangsu, 6.3% for Beijing and 5.8% for Fujian. Columns (2), (3) and (4) show the share for each of the three migrant-sending regions, respectively. Central inner China is the leading migrant-sending region, with 51.1% of the total Hukou registrations. Migrants from Central Interior China to Guangdong Province make up the biggest bilateral flow, with 23.58% of the total. Migrants from Western Interior China account for 27.8% of the total, while those from Northeastern China only represent 4.8%. Table 8 in Appendix C shows the summary statistics on the observable characteristics of non-hukou migrants, while also comparing these characteristics with residence either at destination or at origin. Non-Hukou migrants are on average younger than residents both in the destination province and in the origin province. We also find that they are more likely to be males or to have a low to secondary level of education, while those who are the least or the most educated are underrepresented. 7

Table 1: INTER-PROVINCIAL NON-HUKOU MIGRANTS AS A SHARE OF NATIONAL TOTAL Destination All origins Central Western Northeast Interior Interior (1) (2) (3) (4) Guangdong 35.6% 23.58% 9.79% 0.38% Zhejiang 13.0% 7.48% 4.51% 0.09% Shanghai 9.4% 4.67% 1.40% 0.20% Jiangsu 8.7% 5.52% 2.19% 0.11% Beijing 6.3% 3.53% 0.89% 0.72% Fujian 5.8% 2.96% 2.53% 0.04% All 6 provinces 78.8% 47.7% 21.3% 1.54% All destinations 100% 51.1% 27.8% 4.8% Notes: Each value represents the migration flow of a given origin-destination pair as a share of the national total. Central interior provinces include Hebei, Shanxi, Inner-Mongolian, Henan, Anhui, Hubei, Jiangxi, Guangxi and Hunan; Western interior provinces include Chongqing, Sichuan, Guizhou, Yunnan, Xizang, Shannxi, Gansu, Qinghai, Ningxia, Xinjiang. Northeast provinces include Heilongjiang, Jilin and Liaoning. 2.2 Internal Migrant Employment by Sector We next discuss employment by sector of non-hukou migrants. In Figure 1, we plot the share of total employment represented by non-hukou migrants in the province against the provincial manufacturing employment share, with a 45-degree line (blue dashed line). The size of each circle in the plot is based on the number of migrants each province received. One important message from Figure 1 is that internal migration makes up a significant portion of the overall provincial labor force in coastal provinces and in Beijing. Along the horizontal axis, Shanghai, Beijing and Guangdong appear to the right, with migration labor forces accounting for 35.2%, 28.9% and 25.3% of the overall provincial labor force, respectively. In relation to the vertical axis, most provinces lie above the 45-degree line, and provinces such as Guangdong, Shanghai, Fujian, Zhejiang and Beijing stand out. These observations suggest that internal migrants are disproportionately sorted into the manufacturing sector and play an even stronger role in the manufacturing labor force. The internal migrant employment share is especially striking in Guangdong and Shanghai, where they account for 56% of manufacturing employment in the former and 40% in the latter. A second pattern evident in Figure 1 is that the larger the circle, the further away it is from the 45-degree line. This indicates that in provinces such as Guangdong, Zhejiang and Fujian, which absorb large quantities of internal migrants, migration workers are more inclined to work in manufacturing relative to their overall employment share. In appendix B, we show that this pattern is not as strong among low-skill sectors such as construction, hotel & restaurant, and retail trade, although internal migrants are still slightly disproportionately employed 8

in these sectors. This suggests changes in manufacturing labor demand and changes in sector employment restrictions for internal migrants may also play important roles in the sector sorting of internal migrants..6 Manufacture employment share of migrants.5.4.3.2.1 Hainan Xinjiang Jiangsu Fujian Tianjin Zhejiang Guangdong Beijing Shanghai 0 0.1.2.3.4.5.6 Employment share of migrants Figure 1: MANUFACTURING VS. OVERALL EMPLOYMENT SHARE OF MIGRANTS; CIRCLE SIZE BASED ON THE NUMBER OF MIGRANTS EACH PROVINCE RECEIVED 2.3 Manufactured Exports and Migration Having shown that China s internal migrants are the crucial manufacturing labor force in coastal provinces, we next use China s Population Survey from 1990 and 2005 and China s customs transaction database to relate changes in China s internal migration with China s export growth from 1990 to 2005. 9 In Figure 2 we plot the changes in each province s manufacturing labor force made up of migrants, normalized by each province s manufacturing labor force in 1990, against the changes in manufacturing export volumes, normalized by provincial manufacturing output in 1990. What is evident in Figure 2 is a strong and positive correlation; that is, coastal provinces which face large increases in the migrant labor force in their manufacturing industry relative to initial industrial employment experience faster growth in manufacturing exports relative to initial sector output. The largest change in migrant labor force and export growth is observed in Guangdong Province. The province s manufacturing export volume in 2005 is 6.3 times its sector output in 1990, and the increase in its migrant manufacturing labor force is about 2 9 Data on the province of Hukou registration are not reported in China s 1990 census. We thus measure interprovincial non-hukou migration using the variable of province of residence five years ago. Since Hukou conversion was tightly controlled prior to 1990 (Chan 2009), five-year migration intensity is considered a close proxy to the actual level of inter-provincial non-hukou migrants in 1990. 9

times its sector employment in 1990. Although their growth is small compared to Guangdong, other coastal provinces such as Fujian, Jiangsu, Shanghai and Zhejiang experience an increase in manufacturing exports to about 3-4 times the sector output in 1990. Their migrant manufacturing labor force increases by about 35%-65% compared to their initial sector employment in 1990. 7 Changes in Export, normalized by output 6 5 4 3 2 1 0 Jiangsu Tianjin Hunan Shandong Beijing Liaoning Sichuan Hubei Heilongjiang Jilin Shanxi Shannxi Henan Anhui Xinjiang Jiangxi Gansu Inner Qinghai Guangxi Hebei Yunnan Guizhou Ningxia Mogo Hainan Shanghai Fujian Zhejiang Guangdong 0.5 1 1.5 2 2.5 Changes in manufacture employment, normalized by LFS Figure 2: CHANGES IN PROVINCIAL MIGRANT MANUFACTURING EMPLOYMENT (NORMAL- IZED BY INITIAL SECTOR EMPLOYMENT) VS. CHANGES IN PROVINCIAL MANUFACTURING EX- PORT VOLUME (NORMALIZED BY INITIAL SECTOR OUTPUT) BETWEEN 1990 AND 2005 In a general equilibrium setting, two mechanisms interact with each other to explain the positive correlation documented in Figure 2. First, the boom in labor demand due to China s opening up to trade causes the massive internal migration. That is, as China begins to relax barriers to population movement in the 1990s, the reduction in international trade barriers stimulates demand for manufacturing labor in China s coastal provinces, triggering the massive migration from poorer interior provinces to wealthier coastal provinces, where economic opportunities are better. The second force is that China s internal migrants provide a cheaper manufacturing labor supply, reinforcing China s comparative advantage and promoting China s export growth. We next investigate the timing of China s export and migration surge. We find that the rapid growth of internal migration to coastal provinces took place prior to and during China s rise in the global good market. This suggests that the second force mentioned previously, which is also a driver of China s export growth, is more predominant. In Section 3, we employ a spatial equilibrium model to account for the role of internal migration in China s export growth. 10

2.4 The Timing of Surges in Exports and Migration To shed light on this phenomenon, we provide evidence on the timing of internal migration and export growth in China s coastal provinces. The left panel (a) of Figure 3 plots the time trend of aggregate yearly manufacturing export volume, as well as manufacturing employment of migrant workers for five coastal provinces. 10 The right panel (b) replicates the plot for Guangdong Province only. Each variable is normalized by its initial value. The migration data are taken from China s Population Survey at three points in time: 1990, 2000 and 2005. We obtain more years of data available to plot provincial manufactured export trends, based on China s annual customs transaction data for 1988-1991, 1997 and 2000-2005. 30 50 30 50 Export Growth Rate 20 10 30 Migration Growth Rate Export Growth Rate 20 10 30 Migration Growth Rate 10 10 1 1988 1990 1997 2000 2002 2005 Year 1 1 1988 1990 1997 2000 2002 2005 Year 1 Export Growth Migration Growth Export Growth Migration Growth (a) Coastal Provinces (b) Guangdong Figure 3: GROWTH IN EXPORTS AND MIGRANT MANUFACTURING LABOR FORCE IN COASTAL PROVINCES, 1990-2005 As plotted with the blue dashed lines in both panels, China s exports grew steadily from the late 1980s to 2000, and accelerated after China s WTO accession in the early 2000s. Notably, the rise in China s migrant manufacturing labor force precedes the timing of China s spectacular export growth. In panel (a), migrant manufacturing workers in coastal provinces began to increase during the period 1990-2000, and the growth rate remained steady over the period 2001-2005. The epic scale of migration seems to take place even earlier in China s largest exporting province, Guangdong. As shown in panel (B), the number of migrant manufacturing workers there rises sharply from 1990 to 2000, but the growth pattern slows after 2000. The timing of migration and export growth suggest that China s rapid growth in internal migration to coastal provinces began, if not prior to, no later than the rise in Chinese exports to the global market. The pattern of the growth rate in migration and exports in Figure 3 also indicates that Hukou reform is the main driver of China s internal migration, while trade-induced 10 These five coastal provinces include Guangdong, Shanghai, Fujian, Zhejiang and Jiangsu. 11

internal migration could be small. In the next section, we present a static equilibrium trade model to evaluate the extent to which China s internal migration from 1990 to 2005 reduced prices in China and promoted its export growth. 3 Environment In this section we present a quantitative general equilibrium model with trade and migration. The labor demand side of the model is extended from the multi-sector version of the Melitz- Chaney model (Arkolakis et al., 2017), in which firms choices where to open an establishment within a country. On the labor supply is built on Roy-like model (Lagakos and Waugh, 2013; Hsieh et al., 2013), in which workers choose location and sector by recognizing the costs of migration internally, their productivity in each labor market, and the returns to their productivity in each labor market. In what follows, we denote countries by i and j, subnational regions by m and n, sectors by l and s, and labor groups by g. Let Ω i be the set of regions and Υ i the set of labor groups belonging to country i. In country i, there is a measure M i,s of potential firms in sector s. Every firm produces a unique variety, and decide whether and which subnational region to produce and export. The proofs of our theoretical results are shown in the appendix A. 3.1 Worker Preferences Workers utility function is Cobb-Douglas function over final goods, with expenditure share β s on the final good produced by sector s. U = s C βs s, β s = 1 s 3.2 Production Technology The final goods are nontradable and produced in a perfectly competitive market with varieties from different sources, either locally produced or imported. The final good produced in sector s and region m is a Dixit-Stiglitz aggregator over the input of varieties within the sector, with a constant elasticity of substitution σ. ( Q m,s = ) σ q m,s (ω s ) σ 1 σ 1 σ dωs where Q m,s is the output of the final good in sector s and region m, and q m,s (ω s ) is the input of variety ω s available in sector s. The final goods can be either consumed by households or used 12

as materials in the production of varieties. Varieties produced by firms can be traded. In country i, each firm in sector s can produce a unique differentiated variety ω s. It needs to decide whether and which subnational region m Ω i to open an establishment to serve each destination market. The productivity vector for a potential firm in all subnational regions is randomly drawn from the multivariate Pareto distribution as follows: where Ãi = F i,s (z 1,s,..., z m,s,...) = 1 ( m Ωi A1/(1 ρ) m,s ( [ Am,s (z m,s ) ] ) 1 ρ ϑ 1 1 ρ, m z m,s à 1 ϑ i m Ω i ) 1 ρ. 1/ϑ A m,s captures the average productivity draws in the region m. ρ [0, 1) relates to the correlation between productivity draws across subnational regions. We allow agglomeration effects as A m,s = Ām,sH α m,s, where Ām,s is a constant and H m,s is the total amount of efficiency units in region m and sector s. α determines how sectoral productivity is affected by employment. 11 Firms use efficiency units of labor and materials in the production. If a firm chooses to open an establishment in region m and sector s, its production function follows Cobb Douglas function: y m,s = z m,s h γsh l where h is efficiency units of labor and m sl is the demand for materials from sector l. γ sh is the share of costs spent on labor. γ sl is the share of costs spent on materials from sector l. We impose γ sh + l γsl = 1 to make the production function constant returns to scale. We denote the cost of the input bundle in region m and sector s to be c m,s, so the marginal cost of a firm with productivity draw z m,s is c m,s /z m,s. m γsl sl 3.3 Trade and Prices We view each subnational region as a different destination market, and introduce trade costs in the form of "iceberg" costs. Let d m,n,s denote the amount of production required in sector s of region m to supply one unit of goods in region n in which (d m,n,s 1) units loses in transit. Following the trade literature, we assume d m,m,s = 1 and d m,n,s 1 {m, n, s}. Given our assumptions on the production technology for final goods, firms simply set their prices equal to mark-up σ = σ σ 1 over marginal cost and choose the subnational region with the lowest marginal cost for production. Hence, the price charged in destination market n for 11 There is little agreement on the sources of agglomeration effects in the literature. Studies show that they could come from market potential, population density or (sectoral) employment. See Melo et al. (2009), Combes and Gobillon (2015), Redding (2016), etc. 13

a firm from source country i is: p i,n,s = σmin m Ωi c m,s d m,n,s /z m,s. Firms that export to destination n incur a fixed cost of marketing F n,s > 0 in units of input bundles in the destination. If profits from exporting outweigh the marketing cost, then the firm starts to serve market n. Let E n,s = P n,s Q n,s be total expenditure of region n on the final good in sector s, the maximum price under which profits cover the marketing cost is: ( ) 1/(1 σ) p σcn,s F n,s n,s = P n,s E n,s Assuming σc m,s d m,n,s /Ã1/ϑ i > p n,s for all i, n, s and m Ω i, 12 we can compute the share of expenditure in destination n that comes from source country i (or region m in country i) as: Π i,n,s = M i,sψ i,n,s i M i,sψ i,n,s Π m,n,s = M i,sψ i,n,s ψ m,n,s i M i,sψ i,n,s, if m Ω i where Ψ i,n,s = [ m Ωi (A m,s(c m,s d m,n,s ) ϑ ) 1 1 ρ ] 1 ρ. The share of subnational region m Ωi in country i s total exports to region n is given by ψ m,n,s = (A m,s (c m,s d m,n,s ) ϑ /Ψ i,n,s ) 1 1 ρ. This expression of trade shares resembles the standard gravity equation in Melitz-Chaney model if we separate each country only into one subnational region. Using CES aggregation in the production of final goods, sectoral price index satisfies: P ϑ n,s = ϑ σ ϑ ϑ σ + 1 3.4 Labor Market and Migration ( σcn,s F n,s E n,s ) ϑ σ+1 1 σ M i,s Ψ i,n,s i We group workers based on the province of their Hukou registration and education, indexing labor group by g. We start by living in their Hukou registration province and make a one-time life decision of choosing subnational region m and sector s to work and live domestically. 13 Each worker is inelastically endowed with one unit of time. Workers are heterogeneous in their productivities across subnational regions and sectors. We assume workers from each group g draw a vector of idiosyncratic productivities independently across regions and sectors, 12 This assumption aims to ensure that the firm with lowest productivity draws does not export. This holds empirically, as only a small portion of firms export in practice. 13 We rule out international migration in the model. 14

denoted as {x g,m,s } m Ωi, s, from Fréchet distributions: ( G g ({x g,m,s } m Ωi, s) = exp ) T g,m,s x κ g,m,s. m Ω i, s T 1 κ g,m,s is proportional to the average group productivity draw in region m and sector s. κ captures the dispersion of productivity draw, with a larger κ corresponds to a smaller degree of productivity heterogeneity. As we will see from equilibrium labor flows, κ also captures labor supply elasticity. A worker from labor group g that works in region m and sector s has to suffer a proportional adjustment to income τ g,m,s, following Galle et al. (2015). τ g,m,s captures both amenities and migration frictions. In this setting, workers maximize their utility which is τ g,m,s x g,m,s w m,s /P m. P m is the aggregate price index in region m. Defining the real wage per efficiency unit in region m and sector s to be V m,s = w m,s /P m, the fraction of group g workers that work in region m and sector s equals Λ g,m,s = T g,m,s(τ g,m,s V m,s ) κ n,l T g,n,l(τ g,n,l V n,l ) κ. Conditional on migration choices, the average efficiency units that group g workers supply in region m and sector s are given by: E[ x g,m,s choose m, s] = Γ(1 1 κ ) T 1 κ g,m,s Λ 1 κ g,m,s. 3.5 Equilibrium Conditions In equilibrium, the good market clearing requires: X m,s = n Π m,n,s E n,s where X m,s denotes the total production of region m and sector s. As in Chaney (2008), the total fixed marketing costs spent in destination n and sector s are η = ϑ σ+1 of E σϑ n,s. The total profits made by firms are π = σ 1 of trade flows. Assume all the profits are dissipated by σϑ managers in the production process, which ensures trade balance in the absence of regional transfers. Then E n,s can be written as: E n,s = l γ ls Y n,l + β s [ l γ lh Y n,l + n ] where Y n,l = (1 η) X n,l + ηe n,l is the sum of production costs and marketing costs that take place in region n. The first part on the right side is total expenditure of all sectors on material 15

inputs from sector s. The second part captures consumption of workers on the final good in sector s, where n denotes regional transfers in income. Labor market clearing requires the demand for efficiency units equals the supply in each region m and sector s. γ sh Y m,s /w m,s = g Υ i L g Λ g,m,s E[ x g,m,s choose m, s] Define H g,m,s = L g Λ g,m,s E[ x g,m,s choose m, s] as efficiency units supplied by group g to region m and sector s. Total efficiency units in region m and sector s are H i = g Υ i H g,m,s, which decides agglomeration effects as described before. 3.6 Solving Method Our model involves a number of unknown fundamentals which are difficult to pin down directly. To make the empirical analysis tractable, we follow Dekle et al. (2008) to match the model exactly to the data (the observed equilibrium) in a given year (the year 2005 in our case). We then perform our counterfatual exercises by solving in changes. 14 Denote the proportional change in variable x as x = x x, where x is the counterfactual status, and x is the observed equilibrium status. In this subsection, we briefly present relevant conditions of our model for solving in changes. Proportional changes in trade shares can be expressed as: Π m,n,s = M i,s Ψi,n,s ψm,n,s, m Ω i j x Ω Mj,s Ψj,n,s ψx,n,s j Π x,n,s [ ] 1 ρ, Ψ i,n,s = (Âm,s(ĉ m Ωi m,s dm,n,s ) ϑ ) 1/(1 ρ) Π m,n,s /Π i,n,s ψm,n,s = (Âm,s(ĉ m,s dm,n,s ) ϑ / Ψ i,n,s ) 1/(1 ρ), and the change in unit costs is given by ĉ m,s = ŵm,s l γsh γsl P m,l. Â m,s = Ā m,s Ĥm,s α contains both changes in fundamental productivity Ām,s and changes in agglomeration effects that work through H m,s. Proportional changes in sector-location migration flows can be described as: Λ g,m,s = T g,m,s ( τ g,m,s Vm,s ) κ n,l T g,n,l ( τ g,n,l Vn,l ) κ Λ g,n,l. 14 This method has been frequently applied in the trade literature. See Caliendo and Parro (2015), Galle et al. (2015), Allen et al. (2014) among others. 16

Good market equilibrium now requires: X m,s = n Π m,n,se n,s E n,s = l γ ls Y n,l + β s [ l γ lh Y n,l + n ] where Y n,l = (1 η)x n,s + ηe n,s. Labor market equilibrium can also be expressed in changes: γ sh Y m,s/ŵ m,s = g Υ i ( L g )( Λ g,m,s ) 1 1 κ ( Tg,m,s ) 1 κ wm,s H g,m,s = w m,s H m,s Ĥ m,s Proportional changes in price indices in industry s of region i are P n,s = ( ĉ n,s Ê n,s ) ϑ σ+1 (σ 1)ϑ [ i M i,s Ψi,n,s Π i,n,s ] 1 ϑ, and proportional changes in the aggregate price index in region n is P n = s P βs n,s Define the average welfare for labor group g as W g = m,s Λ g,m,sτ g,m,s w m,s E[ x g,m,s choose m, s]/p m. Its proportional changes can be expressed as Ŵ g = ( m,s T g,m,s ( τ g,m,s Vm,s ) κ Λ g,m,s ) 1/κ Note if we know initial equilibrium conditions, parameters {α, ρ, ϑ, κ, σ} and proportional changes in fundamentals, we can solve changes in relative terms for all the variables after normalization. In our baseline counterfactual exercise, we set M i,s = L g = Ā m,s = T g,m,s = d m,n,s = 1. τ g,m,s is set to ensure that τ g,m,s is the same level as in 1990 within China. 4 Data 4.1 Data Source We match our model to the data in the year 2005, the most recent year for which our data are available. We calibrate our model to 32 countries and a constructed rest of the world. For China, the subnational regions in China are thirty provinces; we group workers based on their Hukou registration province and two education levels. We consider one subnational region 17

(the country "itself") and one aggregate labor group for each other country. We use 29 sectors, including 18 tradable sectors and 11 non-tradable sectors, according to the International Standard Industrial Classification, Rev 3 (ISIC, Rev 3). We provide details on these 29 industries in Table 7 of Appendix C. To perform our counterfactual exercise by solving for the relative changes, we measure: the import expenditure share {Π m,n,s }; the inter-provincial migration rates {Λ g,m,s }; sectoral output {X m,s }; labor income {w m,s L g,m,s }; consumption shares of each sectoral goods, {β s }; the share of intermediate inputs in production, {γ sl }. We also need values of agglomeration parameter α; elasticity of substitution σ; correlation of productivity draws ρ; trade elasticity ϑ; and labor supply elasticity κ. Below we provide a brief description of data sources we use and the calibration of our parameters, with further details provided in the Appendix. Bilateral Trade Flows: We compute bilateral trade flows between China and other countries from 2005 China Customs Transactional Database. This database records detailed geographical information and 8-digit HS classification number of each transaction of imports or exports, which allows us to identify trade flows between each Chinese province and other countries in each sector. Our measures of bilateral trade flows and sectoral output across 30 Chinese provinces are constructed by the China provincial input-output tables in 2007, which is the closest available year to 2005. We adjust these trade flows and output using growth rates of China s sectoral output between 2005 to 2007. Bilateral trade flows across other countries are sourced from STAN Bilateral Trade Database. Sectoral output for other countries, elasticities of intermediate inputs {γ sl } and consumption shares β s are calculated from OECD Inter-country Input-Output Tables. We use these data to calculate the import expenditure share {Π i,j,s } and sectoral output {X i,s }. Labor market: For China, 2005 Population Survey 0.2% sample (2.6 million observations) provides detailed information on the province of Hukou registration, the current province of residence, industries employed, and earnings for each individual. We use these data to construct the inter-provincial migration rates {Λ g,m,s } and labor income {w m,s L g,m,s } for. For other countries, we utilize IPUMS Database and Luxembourg Income Study Database to compute migration probabilities {Λ g,m,s } and labor income {w m,s L g,m,s }. Parameters: Elasticity of substitution σ and trade elasticity ϑ have been conventionally set as σ = 4 (Arkolakis et al., 2017), 15 and ϑ = 4.5 (Simonovska and Waugh, 2014). 16 For migration 15 This is also in line with empirical evidence of Broda and Weinstein (2006) on SITC-3 industries. 16 These parameters imply ϑ σ 1 = 1.5, which is similar to what is found in other studies about China. For example, Handley and Limão (2017) obtains a range from 1.4 to 3.1 using trade policy uncertainty and HS-6 imports from China. 18

elasticity κ, we perform maximum likelihood estimation on the empirical wage distribution over the entire China s working age population who are wage earners. We obtain an estimate of κ = 2.1, which lies in the range of estimates in the literature (Hsieh et al., 2013; Burstein et al., 2015; Tombe et al., 2015). We show this parametric assumption fits the empirical wage distribution decently well in Figure 7 of appendix. Correlation parameter ρ cannot be directly identified from our data. We set ρ = 0 in our baseline calibration, which implies a trade elasticity of 4.5 between subnational regions, similar as the estimates the trade elasticity between Indian districts studied by (Danoldson, 2017). 17 We also experiment with ρ = 0.5 in robustness checks. We set the agglomeration elasticity equals 0.05 following from the literature (Combes and Gobillon, 2015). 18 4.2 Estimating Changes in Policy Barriers We recover changes in policy-related migration barriers from changes in inter-provincial and sector migration flows, analogous to trade literature which calibrates changes in trade costs from changes in trade flows (Head and Ries, 2001). We combine China s population census 1990 with China s Population Survey 2005, to measure changes in province-sector migration rates by each labor group g. 19 The identification of changes in policy-related migration barriers requires two assumptions. First, we assume changes in labor-sector productivity over time are independent from the destination province they work at, that is, Tg,m,s = T g,s. This assumption allows us to attribute variation in changes of migration rates across destination provinces to variation of changes in migration frictions. Second, we impose multiplicative separable assumption on τ g,m,s such that τ g,m,s = τ g,m τ m,s. The first term τ g,m captures non-policy related migration costs that are specific to origin-destination (or education groups). Factors that affect τ g,m may include geographic and transportation costs, culture assimilation, etc. The second term τ m,s captures destination and sector specific policy barriers to migration, and is independent across origin provinces. What this assumption says is that, provincial Hukou restriction can be relaxed more in some sectors than the others, but do not discriminate workers based on their province of Hukou origin. This sep- 17 ϑ In our model 1 ρ captures trade elasticity between subnational regions, while ϑ measures trade elastcities between countries. Danoldson (2017) estimates the trade elasticity between Indian districts, and he finds the trade elasticity has a mean of 3.8 with a standard deviation of 1.2 across commodities, which is quite similar to trade elasticities between countries. 18 Combes and Gobillon (2015) surveys the literature and finds that typical values for this elasticity is between 0.04 to 0.07. 19 Since variable on the province of Hukou registration is unavailable in China s population census 1990, we measure province-sector migration rates basing on a 5-year migration intensity measurement: the mismatch of residential province between 5 years ago and the survey year. 19

arable assumption provides us a convenient way of isolating changes in migration frictions related to geographic and transportation costs which we control them as a function of geographic distance, and changes in migration frictions related to Hukou reform which is the parameters of interests here. Let the hat variable be the ratio between variables in 1990 to those in 2005. Taking the logarithm of proportional changes in sector-region migration flow to have log Λ g,m,s = log T g,s θ θ + log τ g,m + log τ m,s θ + log }{{}}{{}}{{} V { } θ m,s log T g,l ( τ n,l Vn,l ) θ Λ g,n,l }{{} α g,s f(distance) α m,s θ log V n,l m,s }{{} α g (1) The dependent variable is the proportional changes in province-sector migration flow between 1990 to 2005. On the right hand side of equation (1), α g,s is group-sector specific fixeffects which capture the possibility that workers from some groups may become more productivity in working at some sector than workers from other groups. τ g,m captures changes in non-policy related bilateral migration frictions for group g to migrate to m. We assume τ g,m are symmetric between origin and destination, and approximate it as a function of geographic distance, denoted as f(distance). We measure changes in provincial-sector real wages V m,s from the data, and α g captures group specific fix-effects. α m,s are the changes in Hukou policy-related migration costs in each destination province and sector. One may notice that the value of α m,s is identified only up to scale if one estimates equation (1), since an arbitrary sector-specific shift in migration rates could be explained either by changes in labor group productivity at that sector, or by changes in migration friction at that sector, making these two stories observational equivalent to the data. To pin down the levels, we use information on changes in fraction of workers who stay at home province by sectors, while assuming the policy barriers to migrants who stay at home province are unchanged. Denote the home province as m g, changes in the share of homestayers leads to: log Λ g,mg,s = log T g,s θ θ + log τ g,m }{{}}{{} g α g,s f(distance) + log τ m θ }{{ g,s } α mg,s=0 { } log T g,l ( τ n,l Vn,l ) θ Λ g,n,l θ log V n,l mg,s }{{} + log V θ m g,s }{{} As we assume τ mh,s = 1, and hence, α mg,s = 0. Taking difference between equation (1) and (2) to have α g (2) log Λ g,m,s log Λ g,mg,s = f(distance) + α m,s + θ (log V m,s log V mg,s) + e g,m,s In this estimating equation, changes in Hukou policy-related migration costs are identified 20