Specialization Patterns in International Trade

Transcription

1 Specialization Patterns in International Trade Walter Steingress November 16, 2015 Abstract The pattern of specialization is key to understanding how trade affects the production structure of an economy. To measure specialization, I compute concentration indexes for the value of exports and imports and decompose the overall concentration into the extensive product margin (number of products traded) and intensive product margin (value of products traded). Using detailed product-level trade data for 130 countries, I find that exports are more concentrated than imports, specialization occurs mainly in the intensive product margin, and larger and richer have more diversified exports and imports because they trade more products. Based on these facts, I assess the ability of the Eaton-Kortum model, the workhorse model of modern Ricardian trade theory, to account for the observed patterns. The results show that specialization through comparative advantage induced by technological differences can explain the qualitative and quantitative facts. The key determinants of specialization are: the degree of absolute and comparative advantage, the elasticity of substitution and geography. Keywords: Ricardian Trade Theory, Comparative Advantage, Specialization, Import Concentration, Export Concentration I thank Kristian Behrens, Andriana Bellou, Rui Castro, Jonathan Eaton, Stefania Garetto, Ulrich Hounyo, Joseph Kaboski, Raja Kali, Baris Kaymak, Michael Siemer, Ari Van Assche, Silvana Tenreyro and Michael Waugh for their useful comments and suggestions. This paper also benefited greatly from comments by seminar participants at Boston University, Carleton University, Georgetown, the University of Laval, HEC Montreal and the University of Montreal. All errors are my own. Contact: Banque de France, 31 Rue Croix des Petites Champs, Paris 75001, France ( [email protected]). 1

2 1 Introduction The pattern of specialization is at the core of international trade theory. A consequence of international trade is that countries do not need to produce all their goods: instead they can specialize in the production of certain goods in exchange for others. Trade theory offers different explanations of how countries specialize in the number and sales volume of goods. Assessing the empirical relevance of the underlying theory is of vital interest since it not only allows us to evaluate the gains from trade due to specialization but also informs on how trade affects the structure of an economy. For example, a high degree of specialization increases the likelihood that productspecific shocks will have aggregate effects in terms of output volatility and/or an impact on the terms of trade. 1 The main contribution of this paper is to assess the Eaton and Kortum (2002) model s ability to account for the observed specialization patterns. To do so, the paper first documents facts on the pattern of specialization by computing concentration indexes for both, exports and imports. In addition, it decomposes the overall level of concentration into a measure for the extensive and intensive product margin. The extensive product margin indicates the degree of specialization in the number of goods traded. The concentration index for the intensive margin measures specialization in the value of goods traded. Based on the the resulting stylized facts, it evaluates the model on three basic questions about specialization: What explains the level of specialization in exports and imports? Does specialization occur in the intensive or extensive product margin? How does specialization vary with income across countries? The starting point of my analysis is an empirical assessment of cross-country specialization patterns using several measures of concentration for exports and imports as in Cadot et al. (2011). Based on product-level trade data, the results show that countries specialize more in exports than imports and that specialization occurs predominately on the extensive margin for exports and on the intensive margin for imports. This implies that countries export few products and have their foreign sales fairly equally distributed. On the other, countries import a wide range of products but concentrate their expenditure in a small number of products. The difference between export and import concentration is explained by the fact that countries export fewer goods than they import. Focusing on cross-country differences, the results show that countries with higher income have more diversified imports and exports. This diversification is mostly along the extensive margin, i.e. rich economies export and import a wider product range. 2 Having documented the observed specialization pattern, I then simulate the standard Ricar- 1 An illustrating example is the case of Saudi Arabia s and its dependence on oil, which accounts for 90 percent of its export revenue and 55 percent of its GDP. Thus, any change in the price of oil will likely raise volatility and have important macroeconomic consequences. Papers that study the link between concentration of production and volatility are, for example, Jansen (2004), Koren and Tenreyro (2007), di Giovanni and Levchenko (2012) and Haddad et al. (2013). 2 These results are robust to different levels of aggregation and product classification schemes (NAICS, HS, SITC). 2

3 dian trade model developed by Eaton and Kortum (2002) (henceforth EK) and assess its ability to replicate the stylized facts. Following the standard calibration approach in the literature, the model can reproduce the observed specialization pattern qualitatively but not quantitatively. The obtained levels for exports are too high in comparison to the data. The main reason is the underlying productivity distribution. In the simulated model countries export their goods to too many countries in comparison with the data. However, a key benefit of the model is that it sheds light on the underlying determinants of specialization. The principal factors are trade costs together with technology, the elasticity of substitution and the degree of comparative advantage. To replicate the observed cross-country specialization patterns, trade costs have to be asymmetric trade costs as in Waugh (2010) with poor countries facing higher costs to export than rich countries. At this point, it is important to note that the Ricardian model shares with other models of international trade, most notably monopolistic competition models based on Krugman (1980) and Armington models like Anderson and Van Wincoop (2003), the ability to develop quantitative predictions about specialization patterns in the intensive and extensive product margin (see Hummels and Klenow (2005)). However, in these models, products are differentiated by country of origin by assumption. Instead, in EK suppliers from each country potentially supply the same good and a consumer purchases the good from the cheapest supplier only. Thus, the pattern of specialization arises endogenously. As the paper demonstrates, the presence of trade costs and different technologies across countries allows the model to reproduce the facts on the extensive and intensive margin. This paper contributes to the international trade literature on the relationship between income and trade concentration, see Brenton and Newfarmer (2007), Cadot et al. (2011), Koren and Tenreyro (2013) and Papageorgiou and Spatafora (2012). The evidence presented in these papers shows that developing countries have higher levels of export concentration and that these high levels of concentration may lead to more output volatility, Jansen (2004) and Koren and Tenreyro (2007), and less macroeconomic stability, IMF (2014). The results in this paper confirm previous results on exports and shows that the negative relationship between income and trade concentration extends to imports. The novelty of the paper is that it synthesizes the individual facts on export and import concentration along both, the extensive and intensive, margin and shows that the Ricardian model of Eaton-Kortum (EK) can replicate the observed cross-country patterns. As such, the Eaton-Kortum framework provides theoretical guidance for policy recommendations aimed at increasing trade diversification. The analysis also relates to the growing literature in quantifying the importance of Ricardian comparative advantage in explaining trade patterns using the EK framework, (for example, Shikher (2011), Levchenko and Zhang (2011) and Costinot et al. (2012)). These papers specify a multi-sector Ricardian model with both inter- and intra-industry trade in order to derive implications on a sectoral level. In contrast, I abstract from intra-industry trade and attach a sectoral 3

4 interpretation to the continuum of traded goods within the standard Eaton-Kortum framework. Given this notion, the number of traded sectors arises endogenously and is not fixed as in the previous papers. While the standard model has been primarily used to explain bilateral trade flows and trade volume, (see, for example, Eaton and Kortum (2002), Alvarez and Lucas (2007) and Waugh (2010)), I focus on the implications for the pattern of trade and analyze how technology, geography, tastes and comparative advantage induce countries to specialize in narrow sectors. The rest of the paper is organized as follows. Section 2 describes the data and presents the empirical evidence for import and export concentration. Section 3 lays out the theoretical framework. Section 4 describes the calibration and presents the simulation results. Section 5 discusses policy experiments that reduce export concentration. Section 6 shows the robustness of the results and discusses alternative empirical implication based on the Armington assumption. Section 7 concludes. 2 Empirical evidence and data The starting point of my analysis is an empirical assessment of the observed specialization patterns in world trade using detailed product-level trade data. Before describing the data and the empirical evidence, I examine the properties of the concentration measurements used, which form the basis for the qualitative and quantitative tests of the model. 2.1 Concentration measurements I compute two measures of specialization for product level sales, the Gini coefficient and the Theil index. For concreteness, I focus on exports - concentration measures for imports are entirely analogous. The two measurements are defined as follows. Let k index a product among the N products found in the world economy, let r ik be the corresponding export sales revenue, say, in a given country i. The export Gini in this country is defined as: G ix = 1 1 l=1 N x il x il 1 N k=1 N r ik (1) where x il = l k=1 r ik are the cumulative export revenues and indexed in increasing order of size, i.e. r ik < r ik+1. N denotes the total number of products in the world. The Gini coefficient ranges between zero and one. Zero expresses complete diversification, i.e. (1) a country exports all products and (2) each product generates the same amount of revenues. An index of one indicates complete specialization in which case export revenues stem from one product only. The Gini coefficient is independent of scale, follows the principle of population and adheres to the Weak Principle of Transfers, see Cowell (2009). Scale independence (multiplying each observation by a constant does not change the index) 4

5 ensures that cross-country differences in the Gini are solely due to differences in the shape of the export revenues distributions and do not depend on the level or unit of export sales. The Principle of Population (doubling the sample population does not change the index) minimizes the impact of the aggregation bias due to product classification standards in the trade data, in particular proportional changes in the level of disaggregation do not change the concentration index. 3 Differences in the concentration index between countries can stem from two underlying mechanisms. Either a country exports more products relative to another (i.e. differences in the extensive margin) or, given that two countries export the same number of products, the distribution of export revenues of one country is more skewed in comparison to another country (differences in the intensive margin). 4 Because the Gini index does not allow for this distinction, we use the Theil index as an alternative concentration measure. To capture differences in the intensive product margin, we define the intensive Theil as follows: Tix Int = 1 N ix r N ix ik ln r k=1 ix ( rik r ix ) (2) where N ix denotes the number of exported products by country i and r ix = N ix k=1 r ik represents the mean export revenue of country i. The intensive Theil index measures the concentration in export sales by a weighted average of the log distance from the mean. In addition, we define the extensive Theil of country i as follows: T Ext ix ( ) N = ln N ix where N ix denotes the number of products exported and N the total number of products in the world. Thus, the concentration between two countries may simply differ because one country exports more products. To assess the relative importance of each margin in terms of overall concentration, we sum the extensive and the intensive component to get the total Theil index, which is equal to: T ix = T Int ix + T Ext ix = 1 N N k=1 ( ) r ik rik ln R ix R ix where R ix = 1 N N k=1 r ik represents the mean export revenue over all products N. 5 (3) (4) The total Theil index takes the value of zero in the case of complete diversification, i.e. exporting all products (N ix = N) and export revenues are equal for all products, ( R ix = r ik, k). In the case of complete specialization, one product generates all export revenues, (N ix = 1, r ik = r ik, R ik = r ik /N ) 3 In the appendix we show that our results are robust to different levels of aggregation (4 digit versus 6 digit) and different classification systems (SITC versus HS). 4 The Weak Principle of Transfers ensures that if the export revenue of a low selling product increases by the same amount a high selling product decreases its revenue, then the export revenue distribution becomes less skewed and concentration decreases. 5 In the appendix, we show that the sum of the intensive and the extensive component indeed imply the total Theil index given in equation 4. 5

6 and the index takes a value of ln(n). As the Gini index, the Theil is scale independent and adheres to the Principle of Population and the Weak Principle of Transfers. The advantage of the decomposition is that we can focus on differences in the underlying factors that determine overall concentration. This distinction is important from a policy perspective because higher export concentration is associated with higher volatility in export earnings, see Jansen (2004) and Koren and Tenreyro (2007), and may result in lower GDP growth, see Ramey and Ramey (1994). If concentration is mainly determined by the extensive margin, the main policy objective to reduce concentration consists of developing new export products, see Cadot et al. (2011). On the other hand, if the intensive margin is dominant, then the policy implication may be to export existing products to more destination countries in order to balance export revenues across products. Later, the Model section shows the determinants of concentration within the EK framework and discusses potential policy solutions facts about specialization This section establishes 3 facts about specialization. First, exports are more concentrated than imports. Second, the extensive margin determines the gap between export and import concentration. Third, cross-country patterns imply a negative relationship between concentration and income for both, exports and imports. To build the empirical evidence, I use the BACI data set provided by CEPII (Gaulier and Zignago (2009)) and choose the digit HS product classification scheme as the preferred level of disaggregation. I follow Hummels and Klenow (2005) and refer to import flows of the same 6-digit product from different trading partners as different varieties of the same product. I assume that the tradable goods sector corresponds to the manufacturing sector. 6 Using a correspondence table provided by Feenstra et al. (1997), I identify a total of 4,529 tradable products. Data on employment, capital and GDP per capita at international (PPP) prices come from the Penn World Table, see Heston et al. (2009). The baseline sample covers 130 countries representing all regions and all levels of development between 1995 and 2011 (17 years). In total, the sample consists of 2210 observations (country-years). Note that when assessing the model, I assume that a 6 digit HS code corresponds to a product in the model. The model is Ricardian and does not feature intra-industry trade flows within the product. However, in the data some countries export and import the same product. To address the discrepancy between the model and the data, I consider net trade flows instead of total trade flows. That is, I compute net exports for each product-country pair and consider a country as an exporter of a product if net exports are positive and as an importer otherwise. To measure the 6 This is a simplification, but it is reasonable as a first-order approximation because, for all countries in the sample, this represents on average 76 percent of all merchandise imports; the median is 91 percent. 6

7 importance of net trade flows relative to total trade flows, I calculate for each country the share of net trade flows with respect to total trade following the intra-industry measure of Grubel and Lloyd (1975). My results indicate an average intra-industry share of 19 percent. In terms of total value in world trade, the sum of all net trade value represents 66 percent of the total value in world trade. Both findings suggest that net trade flows are a reasonable approximation of total trade flows, particularly for developing countries where the share of intra-industry trade is very small. 7 Table 1: Summary statistics of the average concentration indexes for 130 countries over the period Concentration Gini Theil Exports Theil Imports Exports Imports Extensive Intensive Extensive Intensive Total Margin Margin Margin Margin Total Level Share of total 55% 45% 40% 60% Regression log(gdp per capita) *** *** *** *** *** *** *** [ ] [ ] [0.0414] [0.0383] [0.0539] [0.0381] [0.0166] [0.0410] log(population) *** *** *** *** *** *** [ ] [ ] [0.0240] [0.0250] [0.0348] [0.0299] [0.0196] [0.0325] Observations R-squared Based on net trade flows at the product level, I calculate concentration indexes for each country on all margins for each year and then take the average over the whole sample period. Because the concentration indexes used are independent of scale, the calculation on a year-to-year basis avoids the need to deflate the data. Table 1 summarizes the sample statistics by giving the average yearby-year indexes of the 130 countries. Fact 1: Exports are more concentrated than imports The summary statistics reveal that exports are more concentrated than imports on all margins. In terms of overall concentration, the Gini coefficients of exports is 0.98 and of imports 0.91 and the respective Theil indexes are 4.73 of exports and 2.71 of imports. The results also show that the gap between export and import 7 One reason for the discrepancy between the model and the data is that the level of aggregation in the data is too high. In the appendix I present an alternative approach, where I partition the continuum of goods in the model into product categories (think of HS or SITC codes) using a Poisson process. By aggregating several goods of the model randomly into a product category, two way trade may occur because some of the goods may be exported and some imported. This procedure allows examining specialization patterns based on total trade flows rather than net trade flows. In the rest of the paper, I follow the net trade flow approach. I present the estimation and results of the alternative procedure in the appendix. 7

8 Export Concentration Gini index HS 6 digit Mean of the index from KWT AGO BDI BOL BHR AZE BEN BGD CMR CPV COD CAF ECU ISL ARM BFA CHL CRI DOM KAZ GHA HND NGAETH JAM QAT FJI PRYMWI RWA TKM GMB LAO GNB MDG GEO MLT KGZ MRT NERSTP TCD TTO UZB YEM VENPER MUS TZAUGA SEN NPL SAU URY OMN MAR ALB BLR GTM IRL SLV LKA IRNMDA PHL BIH KEN SYR PAK CYP EGY COL ARG JOR NZL TUN HRV FIN ISR LBN LTUMEX CAN AUS EST LVA VNM PAN RUS UKR MYS SGP ROU ZAF NOR SVK BRA BGRGRC HUN KOR PRT IDN THA SVN SWE TUR CHE JPN HKG POL DNK CZE IND AUT CHN ESPBEL GBR NLD USA FRA GER ITA Import Concentration LCA BRN GAB MDV KNA TJK Export Concentration Total Theil index HS 6 digit Mean of the index from AGO BDI NER CAF COD JAMDV MRT TJK KNA GMB GABBRN TCD GNB TTO GHA KWT CMR CPV RWA BFA BEN ISLETH UGAQATKGZ ARM STP BOL CHL ECU BHR LCA PER NGA TZA PRY SEN YEM MWI UZB LAO MLT TKM CRI SAU MUS JOR KAZ MDGAZE NPL IRN HND LBN FJI GEO SGP CYP ARG VENDOM BGD ALB CAN GTM IRL MDA AUSKEN ISR BIH HKG OMN URY SLV PHL PAN MAR LVA MEX GRC NZL TUN MYS COL BLR HRVHUN FIN NOR EGY LTU LKA EST PAK PRT RUS ZAF VNM UKR SYR SVK ESP BRA ROU KOR POL FRA SVN TUR DNK SWE BGRTHA CZE AUT IDN GBR BEL CHE IND JPN NLDUSA GER CHN ITA Import Concentration (a) Gini coefficient (b) Theil index Export concentration index HS 6 digit Mean of the index from Import concentration index HS 6 digit Mean of the index from Extensive Theil index GNBTCD BDI STPRWA CPV AGO BEN GMB MDV KNA CAF MWI LCA BFA COD ETH GAB MRT TKM YEMUGA BRN NER AZENGA LAOQAT ARM BOL KWT GHA CMR JAMTJK SEN PRY TZA OMN ALB FJIGEO MDGBHRISL KGZ CYP KAZ UZB TTO LBN ECU BIH DOM BGD GRC GTM HND JOR MLT SLV VEN SAU MUS NPL MDA KEN URY PAN IRN CRI CHL HRV LKA LVA EGY LTU MAR PER SYR EST COL NOR TUN BLRNZL HKG AUSARG ROU BGR CAN ISR DNK HUN PAK FINMEX IRL PHL POL PRT VNM RUS AUT SVN SGP SVK UKR TUR BEL CZE IDN CHE ESP BRA MYS NLDGBR SWE ZAF THA FRAKOR USA ITA IND GER CHN JPN Extensive Theil index STP GNB TCD CAF LAO RWA ARM COD BDI TKM GMB CPV KGZMRT NER NPL MWI UZB BFA GEO BEN FJI AZE AGO MDA BIH YEM ALB UGA CMR BLR CHNBHR EST ETH SEN QAT TJK SYR KAZ MDG TZA GHA GER LVA BGRDOM HND ITA JAM JPN KNA LTU MUS PAK KEN IND CYP FRA CZE IDN BOL BGD CRI ISL KOR IRN KWT LCA LKA JOR ARG AUT BRA BEL BRN ESPEGY GTM FIN CAN CHE CHL COL DNK ECU GBR GAB HRV AUS HUN LBN MDV PRYOMN MLT SVN URY UKR VNM SLV TTO USA NGANLD ROU POL MAR PER SWE SVK THA TUR TUN RUS ZAF IRL ISR MYS PHL PRT VEN MEX NOR NZL GRC SGP SAU HKG PAN Intensive Theil index Intensive Theil index (c) Export Concentration (d) Import Concentration Figure 1: Average export versus import concentration for the period 1995 to 2011 for 130 countries concentration is mainly due to the fact that countries export few and import many products. These result also holds for the majorities of countries, see Figures 1(a) and 1(b). In case of the United States, China, Germany, Hong Kong, India, Italy and Panama export and import concentration are almost equal. Fact 2: Extensive product margin is more important for exports and the intensive product margin for imports Turning to the decomposition, Table 1 shows that for exports the main driver of concentration is the extensive product margin with a share of 55 percent. Thus, export concentration is more sensitive to changes in the number of products exported rather than to changes in export revenues. This result is in line with existing evidence on the importance of the extensive product margin on the level of export revenues, see Hummels and Klenow (2005) for product and Eaton et al. (2011) for firm level evidence. On the other hand, for imports the intensive product margin has a higher share with 60 percent in terms of overall concentration. Countries import a wide range of products, thus changes in the import expenditure are the main contributor to changes in overall import concentration. However, there are significant cross country differences in the relative importance of each margin, see 1(c) and 1(d). 8

9 Fact 3: Concentration declines with income and size In terms of cross-country differences, the empirical evidence shows that richer and larger countries diversify more than smaller ones. Table 1 states the estimated coefficients when regressing the log of GDP per capita and the log of population onto the concentration indexes. The Gini as well as the Theil index of both, exports and imports, decrease as GDP per capita and size increase, i.e. smaller and poorer countries are more specialized. This relationship is equally pronounced for exports and for imports, with an R square of 0.63 and 0.66 respectively. Turning the attention to the decomposition reveals that the main driver of the negative relationship is the extensive product margin. The R square for exports is 0.76 and for imports Thus, richer and larger countries are more diversified because they export and import more products. These results differ from Cadot et al. (2011) in two important ways. They find that (1) the intensive Theil dominates the extensive Theil, and (2) the extensive Theil is decreasing in income until GDP per capita reaches a certain level after which the index starts to increase. The underlying reason for the difference in the importance of the extensive versus the intensive margin comes from the fact this study is build on net trade flows rather than gross trade flows as in Cadot et al. (2011). Table 12 in the appendix replicates Table 1 for gross trade flows and the results are similar to Cadot et al. (2011). However, there are key differences: (1) I focus on the cross section rather than on the panel and (2) I include the log of population as an additional regressor in Table 1. The reason I am focusing on the cross section is that the model I am assessing has implications on cross-country differences but no dynamic implications. Including population as a control variable allows me to look at the functional relationship between GDP per capita and concentration by controlling for size effects. Figure 2 shows that this relation is negative log-linear rather than a hump shape as in Cadot et al. (2011). However, in line with Cadot et al. I also find that extensive product margin is at the origin of this relationship, 2(c) and 2(d). In general, the theoretical explanation of why larger countries export more products are based on the Armington assumptions, see Hummels and Klenow (2005) and Koren and Tenreyro (2013). However, in these type of models, by assumption tradable goods are differentiated by location of production and each country is the sole producer of a good. In contrast, in the EK model all countries potentially supply the same good and consumer purchase the good from the cheapest supplier only. The next section presents the relevant parts of the Alvarez and Lucas (2007) extension of the Eaton-Kortum framework and highlights the key features that allows the model to reproduce the stylized facts. 9

10 Export Concentration COD Total Theil index HS 6 digit Mean of the index from AGO BDI NER CAF ETH TCD TJK NGA MRT BFA RWA UGA GHA CMR JAM GMB TZA BEN YEM UZB MWI GNB NPL SEN BGD KGZ BOL ARM ECU PER GAB MDG LAO PRY MDV IRNCHL PHL TKM PAN TTO KWT KAZ SAU CPV AZE LCA BRN KEN HND VNM PAKGEO JOR IND STP GTM MAR DOM KNA VEN ARG QAT MEX LBN ZAF CRI RUS SLV MDA ALB EGY BRA BHR MYS ISL HKG SGP CAN AUS UKRCOL MUS MLT SYR FJI LKABIH IDN CHN TUN THA URY OMN ISR BLR GRC IRL TUR KOR LVA NZL ESP ROU POL HUN FRA CYP PRT FIN GBR NOR JPN USA LTU HRV SVK BGR SWE EST CZE DNKGER BEL CHE AUT SVN ITA NLD Log of GDP per capita coef = , (robust) se = , t = Import Concentration COD GNB Total Theil index HS 6 digit Mean of the index from STP TCD ARM IND CAF NER TJK BEN PHL BDI BFA RWA GMB MRT LAO MWI NPL KGZ GEO TKM YEM AGO UZB CHN CPV AZE MYS MLT USA SEN ETHMDGTZA UGA CMR KEN BGD PAK OMN ITA GER GHA NGA MDA SYR VNM PRY ALB THA HND FJI BIH LCA KAZ IDN SLV BOL JOR LKA EGY MDVUKR DOM KNA JAM LBN IRN ZAF BRA BHR HKG JPN QAT ISR RUS KOR ECU GAB COL GTM MAR PER TUN CRI MUS VEN BGR BLR TTO URY LVA LTUTUR MEX HUN IRL BRN KWT NLD SGP GBR SAU EST ARG SVK ROUCHLPOL HRV SVN PRT CZE GRC BEL CHE NZL ISL FIN ESP DNK SWE AUT FRA AUS CAN NOR CYP Log of GDP per capita coef = , (robust) se = , t = 6.62 PAN (a) Overall concentration of exports (b) Overall concentration of imports Extensive Theil index HS 6 digit Mean of the index from Extensive Theil index HS 6 digit Mean of the index from Export Concentration COD TCD BDI RWA AGO ETH GNB CAF MWI BFA BENNGA UGA NER YEM GMB TZA MRT CPV STP GHA CMR TKM SEN MDG BGD LAO BOL AZE UZB MDV GAB NPL TJK ARM PRYJAM KEN KGZ GEO EGY VEN KAZ IRN SAU KWT SLV MDA HND VNM PAK JOR IND LKA GTM MAR ALB PHLLCA PER DOM BRN QAT KNA LBN COL FJI BIH IDN CHNBRA TTO CHLMEX OMN ARGRUS GRC UKR TUNCRI PAN MUS ZAF BHR BLR ROU TUR AUS THA HRVPOL CYPISL LVA LTU MYS MLT PRT ESP HKG CAN USA NZL BGR HUN ISRKORFRA GBR NOR JPN ESTSVK CZE FIN IRL ITA SGP GER DNK SWE BEL AUT NLD CHE SVN Import Concentration COD GNB STP TCD CAF LAO RWA BDI ARM NER GMB MWI NPLMRT TKM KGZ BFA BEN UZB CPV GEO YEM AGO FJIAZE CHN ETH UGA MDGTZA SEN TJK CMR MDA IND BIH NGA KEN GHA SYR PAK ALB GER KAZ BGD VNM BLR ITA JPN USA IDN UKR HND SLV BOL PRY LKA DOM JOR PHL GTM MAR JAM EGY ECU PER TUN LBN ZAF BGR THA IRN BRA LVA LTU EST COL CRI MUS URY TUR ROU TTO MYS ARG MDV LCA KNAVEN MLT MEX OMN POL BHR QAT RUS KORFRA SVK HUN SVN CZE ESP KWT NLD GBR GAB CHL HRV PAN CYP SAU PRT ISR SWE BEL GRC NZL ISL FIN DNK AUT CHE IRLAUS CAN HKG BRN SGP NOR Log of GDP per capita coef = , (robust) se = , t = Log of GDP per capita coef = , (robust) se = , t = 8.7 (c) Extensive product margin of exports (d) Extensive product margin of imports Intensive Theil index HS 6 digit Mean of the index from Intensive index HS 6 digit Mean of the index from Export Concentration COD TJK PAN NER JAM AGO PER TTO CHL SGP PHL CAF MRTCMR KGZ ECU ZAF MYS GHA LCA ARG MLT NPL MDG GMB TZA SEN BGD HND VNM PAK BOL ARM UZB CRI IRN ISL ISR IRL HKG BRN KWT CAN JOR IND PRY KNA GAB MUS MDV UGANGA BDI KEN LAO MDA GEO UKR LBNTHA KAZ MEX BHR RUS KOR AUS BRA SAU QAT ETHBFASLV YEM SYR FJI GTM URY MAR TUN DOM BLR LVA HUN NZLFIN FRA BIH IDN COL CHN VEN SVK ESP JPN SWE LKAAZE ALB EGY BGR BEN TKM ROU LTU EST TUR CYP PRT OMN POLSVN CZE GRCDNK GBR CHE BELGER NOR USA HRV AUT NLD MWI TCD RWA CPV ITA GNB STP Import Concentration COD PHL TJK IND MLT MYS HKG LCA BRN SGP MDV OMN ISR IRL USA TCD GNB BEN NER ETHMDG BFASLV TZA NGA KEN SEN BGD YEM STPUGA GHA CMR MRT HND AGO PAK VNM ARM PRY KNA GAB BOL GEO JOR EGY LBNTHA SAU ECU LKA GTM BDI MWI GMB SYR MAR JAM AZE IDN UZB PER TUN ALB UKR DOM COL ZAF CHN VEN CRI IRN BRA TTO BHR MEX HUN GRC NZL MUS KAZ CHL CAF RWA NPL KGZ CPV TKMBGR URY ROU LTU ARG TUR SVK RUSPRT ISL KOR FIN LVAEST HRV CZE DNK BEL ITA KWT CHE NLD QAT AUS GBR ESP SWE AUT CAN GER NOR JPN POLSVN FRA MDA BIH BLR CYP LAO FJI PAN Log of GDP per capita coef = , (robust) se = , t = Log of GDP per capita coef = , (robust) se = , t = 3.6 (e) Intensive product margin of exports (f) Intensive product margin of imports Figure 2: Average export and import concentration versus GDP per capita across 130 countries. Note: Each point in the scatter plots represents the difference from the average concentration index and the log of GDP per capita relative to the United States. The data have been adjusted to remove country size (population) effects. 10

11 3 Model The Eaton Kortum model is Ricardian, with a continuum of goods produced under a constantreturns technology. In this paper, I focus on the Alvarez and Lucas (2007) model and include capital as in Waugh (2010). Next, I derive the relevant theoretical predictions on the pattern of trade and evaluate the importance of the key model parameters for import and export specialization. Consider a world economy with I countries, where each country produces tradable intermediate goods as well as non-tradable composite and final goods. Following Alvarez and Lucas (2007), I define x = (x 1,..., x I ) as a vector of technology draws for any given tradable good and refer to it as good x with x R I +. The production of an intermediate good in country i is defined by: q i (x i ) = x θ i [k α i s1 α i ] β q 1 β mi. Technology x i differs between goods and is drawn independently from a common exponential distribution with density φ and a country specific technology parameter λ i, i.e. x i exp(1/λ i ). I denote the interest rate by r i, the wage by w i and the price of the intermediate aggregate good by p m,i. The intermediate good sector is perfectly competitive. Producers of the intermediate good minimize input costs and sell the tradable intermediate good at price p i (x i ) = Bxi θ[rα i w1 α i ] β p 1 β mi. The continuum of intermediate input good x goes into the production of the composite good q i symmetrically with a constant elasticity of substitution (η > 0) [ˆ η/(1 η) q i = q(x) φ(x)dx] 1 1/η. 0 The aggregate output of intermediate good q i can then be allocated at no cost to the production of final goods or can be used as an input in the production of other intermediate goods. Similarly, capital and labor can be used either to produce intermediate or final goods. Finally, consumers draw utility linearly from the final good. All markets are perfectly competitive. Since these features are not central to the implications I derive in this paper, I omit them. I refer interested readers to Alvarez and Lucas (2007) for a full description of the model. 3.1 General equilibrium Once a country opens up to international goods markets, intermediate goods are the only goods traded. Final goods are not traded and capital and labor are immobile between countries. Trading intermediate goods between countries is costly. We define Iceberg transportation costs for good x from country i to country j by κ ij where κ ij < 1 i = j and κ ii = 1 i. As in Alvarez and Lucas (2007), we also consider tariffs. ω ij is the tariff levied by country i on goods imported 11

12 from country j. Incorporating the trade costs, composite good producers in country i will buy the intermediate good x from country j that offers the lowest price p i (x) = B min j [rα j w1 α j κ ij ω ij ] β p 1 β mj x θ j. (5) Equation 5 shows that whether country i specializes in the production of good x depends on the productivity realizations, factor prices and trade costs. If country i does not offer a good at the lowest cost in the local market, the good is imported. Following Alvarez and Lucas, the resulting price index of tradable goods in country i is p mi = (AB) I j=1 [rα j w1 α j κ ij ω ij ] β p 1 β mj 1/θ λ j θ. (6) Next, we calculate the expenditure shares for each country i. Let D ij be the fraction of country i s per capita spending p mi q i on tradables that is spent on goods from country j. Then, we can write total spending of i on goods from j as ˆ p mi q i D ij = p i (x)q i (x)φ(x)dx B ij where B ij defines the set of goods country j attains as a minimum in equation 5. Note that D ij is simply the probability that country j is selling good x in country i at the lowest price and calculated to be D ij = (AB) 1/θ [rα j w1 α j ] β p 1 β mj p mi κ ij ω ij 1/θ λ j. (7) Equation 7 shows that in this model the sensitivity of trade between countries i and j depends on the level of technology λ, trade costs ω, geographic barriers κ and the technological parameter θ (reflecting the heterogeneity of goods in production) and is independent of the elasticity of substitution η. This result is due to the assumption that η is common across countries and does not distort relative good prices across countries. Note also that, by the law of large numbers, the probability that country i imports from country j is identical to the share of goods country i imports from j. In this sense, trade shares respond to costs and geographic barriers on the extensive margin: as a source becomes more expensive or remote it exports/imports a narrower range of goods. It is important to keep in mind that the number of intermediate input industries that enter into the production of the composite good is fixed. Each country uses the whole continuum of intermediate goods to produce composite goods. There are no gains from trade due to an increased number of varieties. Welfare gains are realized through incomplete specialization. Domestic production competes with imports and countries specialize through the reallocation of resources made available by the exit of inefficient domestic producers. 12

13 To close the model, we impose that total payments to foreigners (imports) are equal to total receipts from foreigners (exports) for all countries i L i p mi q i I I D ij ω ij = L j p mj q j D ji ω ji (8) j=1 j=1 The previous equation implies an excess demand system which depends only on wages. Solving this system, describes the equilibrium wage for each country together with the corresponding equilibrium prices and quantities. Next, I describe the calibration and the simulation of the model. 4 Calibration and results To simulate the theoretical model, which assumes an infinite amount of goods, I "discretize" the Fréchet distribution of total factor productivity and calculate the respective trade value for each product x. Regarding the parameters of the model, we need values for α, β, γ, η and θ. For α, β and γ, I use the same values as Alvarez and Lucas (2007): I set the capital share to α = 0.3, the share of value added in the tradable goods sector to β = 0.5 and the share of value added in the production of non-tradable final goods to α = To calibrate the elasticity of substitution (η) and the variance of the productivity distribution (θ), I use more recent estimates from the literature. In line with Simonovska and Waugh (2011) and Bas et al. (2015), I set θ = 0.25 and η = 4. The remaining parameters that need to be calibrated are trade costs and technology. To do so, I follow Waugh (2010) and use the structural log-linear gravity equation, which relates bilateral trade shares to trade costs and model s structural parameters. To derive the relationship, I simply divide each country i s trade share from country j, see equation 7, by country i s home trade share. Taking logs yields I 1 equations for each country i : log ( Dij D ii ) = S j S i + 1 θ log(κ ij) + 1 θ log(ω ij) (9) in which S i represents the structural parameters and is defined as S i = log([ri αw1 α i ] β/θ p (1 β)/θ mi λ i ). In order to estimate the trade costs κ and technology λ implied by equation 9 I use data on bilateral trade shares across 130 countries. I follow Bernard et al. (2003) and calculate the corresponding bilateral trade share matrix as the ratio of total gross imports of country i from country j, M ij, divided by absorption Abs i D ij = M ij Abs i. Absorption is defined as total gross manufacturing output plus total imports, M i, minus total exports, X i. To compute absorption, I use gross manufacturing output data from UNIDO. 8 Combined with trade data from BACI, I get the expenditure share, D ij, which equals the value of the 8 Details are provided in the appendix. 13

14 inputs consumed by country i and imported from country j divided by the total value of inputs in country i. Note that instead of focusing on a particular year, I compute the expenditure share for each year of the period and take the average expenditure share over the sample period. 9 In total, there are only I 2 I informative moments and I 2 parameters of interest. Thus, restrictions to the parameter space are necessary. To create them, I follow Eaton and Kortum (2002) and assume the following functional form of trade costs. log ( κ ij ) = bij + d k + ω ij + ex j + ɛ ij Trade costs are a logarithmic function of distance (d k ) a shared border effect between country i and j (b ij ), a tariff charged by country i to country j and an exporter fixed effect (ex j ). Tariff (ω ij ) represents the weighted average ad valorem tariff rate applied by country i to country j. The distance function is represented by a step function divided into 6 intervals. Intervals are in miles: [0, 375); [375, 750); [750, 1,500); [1,500, 3,000); [3,000, 6,000); and [6,000, maximum]. ɛ ij reflects barriers to trade arising from all other factors and is orthogonal to the regressors. The distance and common border variables are obtained from the comprehensive geography database compiled by CEPII. To recover technology, I use the estimated trade costs, ˆκ, and structural parameters, Ŝ, to compute the implied tradable good prices, ˆp m, by rewriting equation 6 in terms of Ŝ: ( I ˆp mi = (AB) ( ) θ ) 1/θ eŝj ˆκ ij ω ij j=1 From the obtained prices and the estimates Ŝ i, I get the convolution of wages and technology. Given the bilateral trade shares D ij and the balanced trade condition in equation 8, I follow Alvarez and Lucas (2007) and use the relationship between factor payments and total revenue to calculate equilibrium wages. 10 ( 1 w i = (1 s f i )L i ) ( I j=1 ) (1 s f j ) L j w j D ji ω ji F j where s f i is the share of labor in the production of final goods s f i = γ(1 (1 β)f i ) (1 γ)βf i + γ(1 (1 β)f i ) 9 The resulting sample consists of 130 times 129 potential observations if each country trades with all other countries. In our sample the total number of observations is implying a small number of zeros in the bilateral matrix. I conduct a robustness test where I estimate the model with the Poisson estimator proposed by Silva and Tenreyro (2006). The results are similar and are available upon request. 10 Given factor endowments and optimal factor choice, the interest rates equals: r i = α/(1 α)w i (L i /K i ) 14

15 Table 2: Estimation Results Summary Statistics Observations TSS SSR R E E Geographical barriers Barrier Paremeter estimate Standard error % effect on cost [0,375) % [375,740) % [750,1500) % [1500,3000) % [3000,6000) % [6000,max) % Tariff % Shared border % Note: All parameters were estimated by OLS for an estimated parameter ˆb, the implied percentage effect on costs is 100 (e θb 1) with θ = and F i is the fraction of country i spending on tradable goods net of tariff expenses. F i = I D ji ω ji j=1 The obtained equilibrium wages together with tradable good prices, determine the implied technology levels ˆλ for each country given the structural estimates of the gravity equation. Table 2 summarizes the regression outcome of the gravity equation. In terms of fitting bilateral trade flows, I obtain an R 2 of 0.82, which is slightly lower than the R 2 of 0.83 reported by Waugh. The coefficients obtained for trade costs are consistent with the gravity literature, where distance and tariffs are an impediment to trade. The magnitudes of the coefficients reported in Table 2 are similar to those in Eaton and Kortum (2002) and in Waugh (2010), which consider a similar sample of countries without tariffs. The overall size of the trade costs in percentage terms are higher than those reported in Anderson and Van Wincoop (2004) and Waugh (2010). The main reason for the higher cost effect lies in the underlying trade elasticity. Waugh (2010) uses an elasticity of 5.5 whereas this paper uses The trade elasticity equals to 1 θ, see equation 9. 15

16 4.1 Results Next, I feed the model with the estimated trade costs and technology levels. 12 Table 3 shows the mean concentration levels for the simulated countries. The results show that the calibrated model replicates the Fact 1, i.e. countries are more concentrated in exports than in imports on all margins. However, the levels of export concentration are almost twice as high as the ones observed in the data: mean export (import) concentration on the extensive product margin is 5.04 (0.82) compared to 2.60 (1.10) in the data. This implies that, in the simulated model, countries export (import) 0.8 (43.2) percent of the product space compared to 7.4 (33.3) percent in the data. On the hand, the simulated import levels are close to the one observed in the data. Table 3: Summary statistics of the average simulated concentration indexes for 130 countries. Concentration Gini Theil Exports Theil Imports Exports Imports Extensive Intensive Extensive Intensive Total Margin Margin Margin Margin Total Level Share of total 62% 38% 30% 70% Correlation Regression (Simulation θ = 0.25, η = 4)) log(gdppc) *** *** *** *** [0.0010] [0.0023] [0.0305] [0.0500] [0.0524] [0.0190] [0.0108] [0.0244] log(population) *** * *** *** *** *** [0.0016] [0.0034] [0.0455] [0.0747] [0.0783] [0.0283] [0.0162] [0.0364] Observations R-squared Turing the attention to the decomposition shows that consistent with Fact 2, the extensive product margin is more important for exports and the intensive product margin for imports. In quantitative terms, the share of the extensive margin of exports (imports) is 62% (30%) compared to 55% (40%) in the data. This result reflects the previous point that, in the model, countries specialize much more in the extensive margin of exports relative to the intensive margin. On the import side, the opposite is true. The lower part of Table 3 shows the simulated relationship between income per capita, size and trade concentration. The model can replicate Fact 3 (Concentration declines with income and size) for exports but not for imports. Also in line with the empirical evidence is that the negative relationship is driven by the extensive margin with a strikingly similar elasticity between income and export concentration. Next, I plot the corresponding cross-country pattern for simulated and empirically observed concentration levels against the log of GDP. Figure 4 shows the relationship between income and 12 See Table 6 at the end of the paper. 16

17 concentration conditional on population and the mean. The model replicates the empirical pattern with export concentration decreasing as income increases, see 4(a). However, the cross country differences in terms of income and concentration are much larger in the model than in the data. In the model, the difference from the mean level of concentration ranges from -2 to +2, whereas in the model from -4 to +4. The reason for the larger dispersion is the extensive margin, see 4(c). On the import side, the cross country differences in terms of concentration in the calibrated model are similar to the data, see 4(b). With regard to the intensive margin, see 4(f), the results show that, in line with the data, the model predicts no relationship between concentration and income. Overall, the calibrated model is able to replicate the qualitative pattern but produces relatively high levels of concentration compared to the data, particularly for the extensive margin of exports. A potential explanation for the excessive export concentration lies in the underlying productivity distribution. While the model reproduces the bilateral trade volumes, it fails to capture the underlying distribution of trade volumes across products. To shed light on why countries trade in too few products, I follow Haveman and Hummels (2004) and plot the empirical and simulated density of the number of exporters and importers per product. 13 Figure 5 shows the results. The simulated countries export their goods to too many destinations. The assumed productivity distribution generates producers that are so efficient that even firms facing high trade costs can sell their products to numerous destinations around the world. As a consequence, the number of exporting countries per product is small. In the data (in blue) more than a third of the products are exported by 25 or more countries. In the simulation (in red) no product is exported by more than 25 countries. With regard to imports, Figure 5(b) shows that, contrary to exports, the simulated distribution of the number of countries importing a product is closely related to the empirical one. 5 Policy experiments This section considers the following policy experiment: suppose a country wants to diversity its exports. This policy was advocated by many policy institutions in order to reduce the negative impact of volatility on growth, see De Ferranti et al. (2002) and IMF (2014). In order to show the model s implications of such a policy, I consider the free trade equilibrium (κ ij = w ij = 1, i, j), which is simple and can be solved by hand. Later, I will use the calibrated model to calculate precisely the policy effects for each country. Within the Alvarez and Lucas (2007), there are basically two ways to decrease export concentration. The first one is an increase in the level of technology (λ), for example through R&D investment, and the second policy is by increasing competitiveness, for example through an exchange rate depreciation. I consider a change in the level of technology as a long term effect, while a change in the exchange rate as a short run effect. 13 To get the empirical distribution of the number of exporters and importers per product, I count for each HS code the number of countries that net export or net import the product. Similarly, the model implied distribution represents the number of exporters and importers for each simulated product. 17

18 5.1 Short run (partial equilibrium) For the moment, I consider the short run equilibrium, where wages cannot adjust. Note that the Theil index of exports is defined as the sum of the extensive and the intensive component, i.e. T ix = T Ext ix + Tix Int. The extensive component is given by T Ext ix and the intensive component is given by T Int ix = 1 N ix = ln N i k=1 ( ) Nix N r ik ln r ix where r ik are export revenues of product k, r ix = N ix k=1 r ik represents the mean export revenue of country i, N ix denotes the number of exported products by country i and N is the total number of products in the world. In the free trade equilibrium, all goods a country produces are exported. Thus, the total value of exports is ( rik r ix ) X i = L j p mj q j D ji j =i and, given that the share of goods produced equals the number of goods produced, we can write the share of goods exported by country i as, n X i : n X i = D ii. In the special case of the free trade equilibrium, every country exports the same goods to all countries. Thus, the distribution of the export revenues is identical to the import expenditure distribution. In the appendix I show that the import and export concentration in the free trade equilibrium are Frechet distributed. The resulting Theil index of the intensive margin equals T Int ix = (ˆ 1 ( Γ(1 1 α ) ln u ( 1/α)) ) u ( 1/α) e u du ln (Γ(1 1α ) ) 0 (10) Concentration on the intensive margin is completely determined by the shape parameter α = 1/(θ(η 1)), which is a function of the elasticity of substitution η and the degree of comparative advantage θ. 14 To compute the Theil of the extensive component of exports, I use equation 7 and impose the free trade equilibrium to get the number of goods exported as a function of wages and 14 The integral in equation 10 cannot be solved analytically. To compute the exact Theil index implied by the shape parameter α, I approximate the integral numerically via the Gauss Laguerre procedure. The derivation can be found in the appendix. 18

19 technology: ix = ln ni X = ln T Ext k αβ θ i k=1 N k αβ θ k w β θ i λ i β w θ k λ k Suppose now that a country wants to diversify its exports and decrease the Theil by 0.1 units. In the free trade equilibrium, the only way to achieve this is goal is by diversifying on the extensive margin. The diversification of exports can be achieved by reducing the local unit labor costs relative to other countries, for example by depreciating the currency (in this framework, this is equal to subsidizing production). To show how an exchange rate depreciation helps to increase exports, I express the unit costs of production in local currency by introducing country i s exchange rate vis-a-vis the US dollar 15, s i, into the free trade version of the model. In this case, the number of goods exported is given by the ratio of local unit costs of production relative to the average unit costs in the importing country and equals ix = ln ni X = ln T Ext s 1 θ i k αβ θ i N k=1 s 1 θ k w β θ i λ i k αβ θ k β w θ k λ k The resulting elasticity on the extensive margin of exports equals to T Ext ix ln s i = ln nx i ln s i = 1 θ (1 nx i ) Hence, the percentage depreciation in the exchange rate needed to increase the share of goods exported by 1 percent is ln s i = ( θ/(1 n X i ) ). The smaller the share of products exported (1 n X i ), the less depreciation is needed to diversify exports Long run (general equilibrium) The general equilibrium version of the policy experiment accounts for the endogenous response of wages. In the free trade version, the equilibrium wage is given by which can be solved to L i w i = w i = k I L j w j D ji. j=1 αβ ( ) θ (β+θ) λi (β+θ) i L i Substituting back into D ii we get the elasticity in the general equilibrium framework. 15 I chose the US dollar as international currency. Of course, I could have chosen any other currency as the common currency denominator. 19

20 (β+θ) ix = ln ni X i L = ln αβ k=1 N k T Ext and the elasticity equals to αβ β θ k (β+θ) (β+θ) i λi β θ (β+θ) (β+θ) (β+θ) k Lk λk T Ext ix ln λ i = ln nx i ln λ i = θ (β + θ) (1 nx i ) Thus, within the general equilibrium, we will have a weaker response to a technology shock in comparison to the partial equilibrium, where the elasticity is TExt ix ln λ i = (1 D ii ). The reason is that an increase in technology will also increase wages, which will partly reduce the pro-competitive effects on the export markets. If we consider now the case of an exchange rate depreciation, the market clearing conditions in US dollars becomes whereas the trade share equation equals I L i w i L j w j D ji = s i, s j=1 j ( [k D ji = (AB) 1/θ α i w i ] β p 1 β mi p mj s j s i ) 1/θ λ i. p mj s i Note that in the free trade equilibrium, prices in US dollars equalize across countries, = p m and the equilibrium wage in local currency becomes: p mi s i = w i = k αβ ( ) β (β+θ) λi (β+θ) i si L i Changes in the exchange rate lead to a one to one change in the wage. Plugging the obtained wage back into the trade share matrix (D ij ), we observe that changes in the exchange rate have no effects on trade. D ji = p β θ m k αβ (β+θ) i Note that the described effects in the partial and the general equilibrium only hold in the free trade equilibrium. Once we introduce trade costs, exchange rate changes will have real effects in the general equilibrium framework because the gain in competitiveness from an exchange rate depreciation is not uniform across countries. Still, the free trade case highlights that policies aiming to diversify in exports (1) depend on the initial share of goods exported and (2) are more effective under the assumption of rigid prices (short run) than in the case of full price flexibility (long run). Next, I show the simulation elasticities for both, the general as well as the partial equilib- λ θ β+θ i L β β+θ i 20

21 rium, in the full model. In particular, for each country in the sample I conduct two experiments: (1) I decrease the exporting trade cost by 10 percent and (2) I increase the level of technology by 10 percent. To obtain the country specific elasticities, I divide the obtained change in the export concentration index by the log Table 4: Simulation: Policy experiments in reducing export concentration Technology PE Technology GE Trade Costs PE Trade Costs GE (1) (2) (3) (4) log(share of products exported) *** *** *** *** [0.0176] [ ] [0.0416] [0.0103] Constant *** *** *** *** [0.0963] [0.0333] [0.234] [0.0554] Observations R-squared Table 4 presents the resulting simulated elasticity of reducing export concentration. We observe that within the partial equilibrium, the simulated elasticity is much higher than in the general equilibrium. In the case of increasing technology, the average elasticity is 0.84 in the partial compared to 0.16 in the general equilibrium. As expected, reducing trade costs has a much stronger effect. The average elasticity is 3.6 in the partial compared to 0.79 in the general equilibrium. In addition, Table 4 also shows that the higher the share of goods exported, the smaller the export concentration elasticity. This confirms the prediction of the free trade version of the model. It is easier for poorer and less open economies to diversify than for rich economies that export already a lot of products. 6 Robustness 6.1 Alternative classification schemes This section addresses concerns about robustness of the observed empirical concentration indexes. In particular, the level of disaggregation as well as the chosen classification scheme may affect the empirical concentration measures and the decomposition of the intensive and extensive margins. For this reason, I re-calculated the concentration indexes for both margins using (1) 5-digit SITC product codes and (2) 6-digit NAICS codes instead of the 6-digit HS codes. The advantage of the NAICS and SITC classification system is that the products are grouped according their economic function as well as their material and physical properties, rather than for tariff purposes as in the HS system. Table 5 shows the calculated concentration indexes based on SITC and NAICS classification and their correlation with respect to HS based concentration indexes. The qualitative estimates for all classification are very similar: exports are more concentrated than im- 21

22 ports; concentration is driven by the extensive margin for exports and by the intensive margin for imports; and in terms of cross-country evidence, larger countries import and export more goods. Strikingly, the L pattern of the extensive margin also appears when the SITC and NAICS classifications are used. Differences between the various classification schemes appear in the levels of import concentration. The reason for this is that the SITC and NAICS classifications comprise a much smaller number of codes than the HS system (2,442 and 460 codes respectively for SITC and NAICS, versus 4,529 for the HS system). Overall, however, the high correlation between the different classification standards shows that the results are robust to the classification system. 6.2 Intra-industry trade In this section I address the discrepancy of the product space in the data and the model caused by intra-industry trade. In the main part of the paper I established a correspondence between the model and the data by netting out within product trade. This approach leaves out valuable information and may bias the results. In an alternative approach, I deal with intra-industry trade by developing a measurement device that enables the model to characterize intra- and interindustry trade. The basic idea is that, in reality, the true state of the world is indeed Ricardian, i.e. varieties are in fact products, but the data are not sufficiently disaggregated to capture the true number of products. Instead, these Ricardian products are aggregated into sectors according to a classification scheme, i.e. HS codes. The suggested procedure converts the measurement of product units in the model to product units in the data and allows us to examine gross trade flows. Because the classification scheme is unobserved, I assume that varieties are randomly assigned an HS code following a Poisson process. Using the structure of the model, I can then estimate the Poisson parameter and characterize the measurement device. I obtain a value of 0.94 for the Poisson parameter implying that, on average, one Ricardian product is equivalent to one HS product category. Based on this result, I group simulated Ricardian products randomly into artificial HS codes and calculate the implied concentration indexes. The results, presented in detail in the appendix, show that this approach produces similar results to the net trade flow approach. 6.3 Implications of alternative trade models Finally, I want to compare my analysis to alternative trade models, in particular, to monopolistic competition models based on Krugman (1980), and Armington models based on Anderson and Van Wincoop (2003). The key difference with respect to the Ricardian model is that in the alternative models, tradable goods are differentiated by location of production. Applying this definition of the product space to the data implies that each country is the sole producer/exporter of an HS code and demands all country-product combinations. Hence, the number of potential goods exported is 4,529 and the number of potential goods imported is 4,529 times 129 trading partners. Table 5 presents the corresponding concentration indexes. The results show that, contrary to the findings of my model, countries are more specialized in imports than in exports and the ex- 22

23 Table 5: Mean concentration indexes for gross trade flows based on the Armington assumption: 130 countries Gini Theil Exports (X) Theil Imports (M) Extensive Intensive Extensive Intensive Exports Imports Total Margin Margin Margin Margin Total Mean index (HS 6 digit) % share of overall concentration % 59% 56% 44% tensive margin drives the import concentration. The extensive Theil of imports implies that the average country imports only products (around 3 percent) out of the 4529 times 129 available products. This suggests that the empirical implications used to evaluate a model depend on the definition of a product, i.e. Armington assumption versus perfect substitutes as in the classical models of comparative advantage. While it is certainly possible to produce the results in Table 5 using a model based on the Armington assumption, the underlying mechanism to generate specialization will be very different. 16 In this paper, the analysis is based on the assumption that foreign varieties are perfect substitutes for domestic ones. One motivating observation is that the Grubel and Lloyd (1975) index of 0.19 indicates that the majority of the trade flows are inter-industry (81 percent) rather than intra-industry. However, I cannot reject these alternative hypotheses for the observed concentration patterns and would like to pursue them in future research. 7 Conclusions I have argued that examining export and import concentration in combination and decomposing them into a measure of extensive and intensive product margin concentration provides new quantitative and qualitative evidence on specialization patterns in world trade. Based on detailed trade data, the calculations show that exports are more concentrated than imports on all margins and that specialization is driven by the extensive product margin for exports and by the intensive product margin for imports. The extensive product margin explains the gap between export and import concentration and drives specialization differences across countries. Larger economies diversify more because they export and import more products. Furthermore, I show that the Eaton Kortum model is consistent with the observed patterns and partly replicates the stylized facts as well as the cross-country differences qualitatively but not quantitatively. Overall, my results stress the importance of geography and absolute as well as comparative advantage in determining the 16 For example by introducing fixed trade costs (see Romer (1994)) or declining marginal utility of varieties (see Ottaviano et al. (2002)) 23

24 pattern of specialization. By looking through the lenses of export and import concentration, this paper analyses how openness to trade changes the production structure of an economy and how these changes relate to income. My results show that the relationship between income and concentration is primarily driven by the extensive margin. This relationship has important macroeconomic policy implications. Specialization increases a country s exposure to shocks specific to the sectors on which its economy concentrates. As a result, the likelihood that product-specific shocks will have aggregate effects in terms of output volatility and/or the terms of trade increases with trade openness. A country can mitigate these negative shocks by diversifying exports through the reduction of unit costs of production (though R&D investment or currency depreciation) relative to the rest of the world. 24

25 8 Figures Export Concentration Gini index HS 6 digit Mean of the index from HKG VNM LBN EST STP GMB PAN SVN CYP ETH BGR BEN CRI AZE ALBCPV ARM BDI ECU BIHBLRBOL BFA AGO AUS BGD BHR BEL AUT CHL COL DOM CAF BRA CMR BRN COD CZEGRC GHA HRVHUN GEOEGY GTM HND GAB GNB FJI CHE DNK IND GBR IRL NOR MLT SVK LKA LTU POL LVA LAO PAK NER URYUKR MRT ISL RWA MDA LCA OMN SYR JAM NGA MAR ROU QAT SEN KNA TUN PRY PHL JOR MDG MEX MUS KEN KAZ NZL MDV KGZ NPL KWT IRN ISRPER MWI NLD PRT RUS SAU TUR SLV UGA TZATTO ZAFYEM AGO VEN UZB TKM BOL BHRAZE BDI SGPESP SWE FIN BGD CMRBRN CPVBENCAF ECU ETH GAB ISL KWT COD ARM BFA ARG MYS THA CHL IDN CRI DOM HNDHKG GHA FJI JAM MDV KNA TTO KAZ NGA LCA GEO KGZGMB LAO GNB MAR ALB FRA CAN KOR ITA BLR BIH EGY GTM IRLIRN LKAMDA MDG QAT PRY MWI RWA TKM MLT MRT NERSTP TCD TJK PAN SAU PERMUS TZAUGA SEN UZB YEM VENURY OMN NPL SLV COL ARGJOR KEN PHL SYR PAK CYP NZL CHN HRV FIN ISR LBN LTUMEX GER AUS EST LVA MYS ROU RUS TUNVNM UKR SGP ZAF SVK NOR BRA USA BGRGRC CAN IDN HUN KOR JPN PRT THA SVN SWE TUR CHE JPN POL DNK CZE IND AUT CHN ESP BEL GBR NLD FRA USA ITA GER Simulation Data Import Concentration Export Concentration Total Theil index HS 6 digit Mean of the index from FJI GMB SEN ETH GHA ALB CPV LCA MDG ARM GAB KEN BDI MDV TJK AGO AZE BIH BEN BFA CAF CMRCOD GEO JAM KNA GNB KGZ BRN LBN CYP ISL MRT NGA MUS MWI RWA SLV TZA NPL UZBTKM TCD NER MDA OMN HND UGA TTO BOL BHR DOM PAN LKA MLT LAO SYR JOR ECU YEM CRI KAZ BGD IRN EST STP LVA EGY GTM COL BGR HRV MAR PRY TUN PER LTU QAT NZL PHL SAU ZAF VEN ROU BLR AGO KWT ISR VNM GRC PAKAUS SVN SVK URY CHL HUN UKR TURBDI CAF COD JAMDV HKG NOR GAB BRN KNA LCA IND BRA PRT MEX CZE POL CMR RUS CPV BFA ARG BEN ARM CHL BEL DNK ECU IRL SGP BOL ETH GHA GMB MRT NER TJK GNB ISL KWTKGZ NGA AUT FIN MYS NLD CHE GBR ESP SWE THA BHR LAO CRI IDN ARG AUS CAN FRA KOR ITA ALB AZE CYPDOM BGD FJI GEO CAN GTM HND MLT MUS IRN JOR LBN KAZ IRL MAR ISR BIH HKG BLR GER CHN USA HRVHUN COL FIN GRC EGY LVA FRA ESP BGR BRA JPN CZE DNK EST KEN MDA MDG MWI PAN PERPRY QAT TCD TTO RWA TZA UGA SAU SEN STP YEM UZB TKM NPL VEN SGP MEX NOR NZL SLV OMN PHL LTU LKA PAK MYS ROU POL PRT TUN URY RUS ZAF SVK UKR VNM SYR SVN TUR SWE KOR AUT IDN GBR BEL CHE THA JPN IND NLD GER CHN USA ITA Simulation Data Import Concentration (a) Gini coefficient (b) Theil index Extensive Theil index Export concentration index HS 6 digit Mean of the index from FJI GMB GNBCPV KNA LCA CAF MDV TCD BRN NPLAGO TJK BEN ISL MRT RWA TKM STP MWI KGZ BDI ALB LAO BIH BFAETH BOL YEM NER MDG COD JAM MDA ARM PRY MLT AZEUZB MUS GAB BGDOMN UGA TTO GEO TZA TUN ECU CYP CMR GHA LKA LBN SLV SEN HRVMAR JOR EST BHR KAZ HND LTU LVA SYR GNB NGA RWA TCD STP BDI KEN CPV SVN URY AGO BLR BGRGTM PER EGY BEN NZL PAKPHL VEN GMB MDV PAN BFA CAF QAT MWI ETH KNA COD TKM GAB MRT SVK YEM UGA BRN LCANER CHL ROU KWT COL CRI IRN DNKCZE GRC ZAF AZE NGA LAOKWT BOL ARM GHA CMR HKG HUN CHE BEL PRT NLD IRL NOR JAM VNM ALB FJI BHR CYP GEO UKR AUS BIH DOM BGDCRI ECU AUT POL GRC GTM HND ISL KGZ LBN MDG SEN QAT PRY TZA OMNKAZ UZBTTO TJK SLV VENSAU JOR HRV LKA CAN KEN IRN GBRCHL EGY LTU IDN FIN ISR KOR MEX SAU EST COL MAR MDAMUS NPL MLT URY BLR LVA PAN SYR PER NOR NZL AUS HKG ARG SGPRUS INDTUR BGR CAN FRA MYSWE AUT BEL CZE DNKHUN FIN MEX ISR IRL POL PRT TUN ROU SVN SVK UKR PAK VNM PHL SGP IDN TUR CHE ESP BRA RUS ITA ESP THA ARGBRA GBR MYS NLD SWE THA ZAF FRA ITAUSA JPN GER GER IND KOR CHN JPN CHN USA Simulation Data Extensive Theil index Import concentration index HS 6 digit Mean of the index from STP TKM GNB TCD UZB BRN COD CAF LAO RWA ARM MWI BDI TKM GMB COD VEN PER TJK FJI TCD CPV KGZMRT NERBRA CMR NPL MWI UZB BFA GEO BEN ARG KWT ISR IRN CAF FJI DOM COL AUS BOL BGD BHR MDA BIH AZE AGO YEM TTO YEM ALB AGO ECU CHL KGZ NPL BFA CHN IND BDI BLR JPN UGA CMRNZL MDV UGA TZA ZAF KAZ KEN USA BIH BLR CHN BHR EST ETH KAZ MDG SEN QAT TJK SYR TZA GHA GERMEX MUS SLV HND GAB CPV ARM GNB LVA IDN PHL MDG SAU TUR GTM BGR LTU MUS DOM KNA CRI JAM JOR QAT EGY RUS ALB LCA OMN MDA BEN FIN AZE RWA GEO LAO ISL NGA HUN SYR SEN UKR ROU TUN CYP CZE JAM HND BOL JPN PAK KEN ITAKNA IND FRAIDN CRI LKALBN ISL KOR IRN BGD KWT JOR URY MAR LVA PAK MRT ESP GER ETH POL GHA HRV KOR LKA CYP ITA LTU NER ARG GTM AUT FIN BRA BEL SGP SWE BGR ESP CZE MLT GRC BRN CHL COL DNK ECU CAN EGY GBR LCA MLT NGA PRY HRV GAB AUS HUN CHE MDV OMN SVN URY UKR VNM SLV TTO NLD ROURUS POL MAR PER SWE SVKTHA USA TUR TUN ZAF MYS GRC CAN PAN EST NOR SVK PRT LBN STP AUT IRL ISR MYS PHL PRT VEN MEX NOR NZL VNM SAU GBR GMB SVN SGP DNK IRL BEL CHE FRA HKG HKG NLD PAN Simulation Data Intensive Theil index Intensive Theil index (c) Export Concentration (d) Import Concentration Figure 3: Average export versus import concentration for the period 1995 to 2011 for 130 countries 25

26 Export Concentration Total Theil index HS 6 digit Mean of the index from NGA ETH Simulation Data COD AGO KEN AGO TZA GHA SEN MDG MWI NPL UZB COD CMR TJK BDI BGD NER UGABFA CAF IRN BDI RWA TCD NER AZE KGZ ARMALB TKM CAF YEM GMB SLV ETH SYR GAB LKA GEO EGYFJI NGA TCD TJK MRT DOM GHA CMR BEN BOL HND JAM ECU BFA BIH RWA UGA JAM TZA MRT GNB MDA MAR MUS CPV INDCOL KAZ BEN GMB LAO LBN JOR PERGNB ZAF YEM MDV OMN PAK GTM PHLVEN BGDKGZ UZB PER MWI BOL ARM ECUIRN GAB CHL MDG NPL SEN LAOPHL PRY MDV KWT PAN CRI TTO TKM KAZ PAN VNM SAU CPVAZELCA TTO SAU BRN BRN ARG QAT LCA TUNKENPAK VNM IND HND GEO JORVEN PRY ROU BRA BHR BGR CYP BRA CRI MEX DOM LBNKNA RUS MAR BHRCAN GTM EGY HKG ISL AUS UKRBLR LVA TUR MDA HRV AUS CHN ALB ISL COL MUS MYSMLT SGP SLV STP ZAF UKR ISR SYR LKA IDN NZL ISR OMN KNA CHL MEX RUS FJI BIHBLR GRC IRL THA URY TUNTURKOR GRC LTU ESTMLT LVA HUN CYP NZL ESP FRA JPN USA ROU POLPRT KWT FIN GBR NOR LTU HRV SVK IDN CHN QAT ARG POL HUN SVK BGREST CZEDNK GER SWE BEL CHE URY AUT STP THA ITA NLD SVN USA PRT SVN HKG ESP GBR CZE NOR MYS ITA NLD AUT BEL FRA SWE SGP KOR CAN GER DNK JPN IRL FIN CHE Log of GDP per capita Import Concentration Total Theil index HS 6 digit Mean of the index from TKM GNB TCD Simulation STP PAN UZB COD BRN COD MWI TCD TJK IND CMR FJI PER VEN ARM CAF CAF NER TJK BEN BRA PHL IRN KWT ISR BDIBFA RWAMRT GMB LAOARG GEOTKM BHR BGD DOMMWI NPL KGZ AGOUZB AGO BOL COL BFA NPLYEM TTO YEM CHN KGZ MDV BIH BLRAUS AZE MYSMLT USA CPV JPN BDI UGA GNB TZA ECU ZAF CHL SEN KENCPV ARM GAB KAZ GTM AZE ALB HND IND MUS CMR SLV NZL MDG KNA CHN BEN LCA EGY CRI GEO JAMAR JOR USA ITA GER ETH PRY SAU BGD TUN SEN MDA IDN MEX NGA MRT PHL OMNROU RUS TUR BHR JPN MDG TZA UGA PAK OMN QAT KEN GHA HKG NGAQATPRY THA ALB BIH RWA SYR UKR ISL FIN NER URY GHA BGRCYP GRC HRV LAO LVA LTU MLT HUN BRA KAZ VNM SYR IDN BRN EGY FJI IRN ISR MDA LCA GBR HND BEL CHE PAK THABOL LKADOM MDV LBN POL ETH ESP CZEITA GMB STP LKA NOR EST GER LBN PAN PRT HUN IRL JOR KNA RUS KORKWT NLD SGP UKR ZAF SLV MEX JAM ECU COL SVK GAB SWE BGR BLR TTO SAU GTM CRI SVN ARG LTU ESTESP GRC LVACZEDNK SGP FRA MARPER TUN VEN MUS TUR SVKNZL AUS CAN FRA DNK KOR IRL AUT CAN MYS ROU GBR CHL POLPRT FIN SWE URY NLD ISL NOR BEL CHE VNM HRV HKG SVN CYP Log of GDP per capita Data (a) Overall concentration of exports (b) Overall concentration of imports Export Concentration Extensive Theil index HS 6 digit Mean of the index from NPL ETH AGO TCD Simulation Data COD CAF GMB MWI TJK BDI BEN BFA BGD RWA FJI GNB MDG YEM NER TKM NGA KGZ UZB TCD BOL CPV LAO AGO UGA TZA MRT ALB COD BDI RWA ETH BIH GNB MDV BFA BEN GHA CMR AZE PAK CAF MWI UGA ARMEGY KEN LCA LKAJAM MDA MAR NER NGA YEM PRY ECUDOM PHL TZA GMB CPV IND STP MRT GHA CMR BRN TKM SEN SYR GEO MDG TUN BGD SEN LAOBOL UZB AZE GAB IRN KAZ PER ARM JAM SLV VEN ECUIRN KAZ KEN KGZ MDV NPLTJK PRY VEN SAU EGY KWT KNA GAB HNDVNM ISL GTM LBN COL MUS JOR OMN PAKGEO ZAF IND HND MAR ALB BRN LKA GTM JOR PER DOM QAT SLV VNM CHN CHN COL LCA LBN ARG STP IDN HRV BIHBRA MEX CHL IDN KNA OMN SYR PHL MDAFJI RUS TTO UKRCRI BHR GRC AUS BGR ROU TTO TUNTUR ROU PAN BLR MUS URY POL HRVCYP CAN UKRUS CHL MEX BLR ESP BRALTU MLT BGR FRA GBR HKG ISL USA ZAF THA MYS LVA LTU HUN MLT PRT KOR NZL ISR JPN NOR EST SVK URY CYP LVA CZEDNK ITA FIN GER AUT PAN GRC TUR GBR EST AUS NZL POL BEL IRL SGP NLD CHE USA SVN SWE HUN SVK BHR CAN SVN CZE KOR JPN CRI SAUTHA PRT HKG FRA NLD GER BEL ESP ITA ARG KWT DNK MYS NORAUTCHE QAT ISR IRL SWE FIN SGP Log of GDP per capita Import Concentration Extensive Theil index HS 6 digit Mean of the index from COD UZB TKM GNB TCD STP Simulation BRN MWI COD CAF TCD TJK LAO FJI PER VEN RWA CMR BDI ARM NER GMB TKM CAF BRA MRT KGZ IRN MWI NPL BFA BEN ARG KWT UZB ISR CPV GEO DOM BHR BGD BOL COL AGO NPLYEM TTO YEM AGO AUS AZE CHN BFA KGZ MDV BDI ECU BIHCHL BLR FJI UGATZA JPN GNB CPV GAB KAZ ETHUGA ZAFCMR IND MDA BIH NZL KEN ALB GER ARM SLV KNAHND CHN AZE ALB GTM IND MUSMDG TZASEN TJK MDG BEN LCA PRY CRI GEO EGY JAM IDN JOR SAU MEX PHL NGA MRT MDA MAROMN QAT ROU RUS GHA PAK SYR RWA SEN TUN TUR USA KAZ JPN BGD BLR ITA NGA KENVNM USA IDNUKR SYR UKR ISL FIN NER URY LAO THA BOL LVA HRV HUN BRA BGR BHR QAT HND LKA GHA BGRCYP LTU ETH LKA PAK EGY DOM THA IRNESTKOR FRA SLV PHL PRY GBR GTM JOR JAM LVA LTU POL ECU COLARG MLTCRI CZE ESP KWT MAR LBN MUS MYS MEX OMNNLD PER ZAFTUR ROU URY RUS POL TUN TTO ESP CZE HUN MLT ITA GMB STP EST NORGER AUT KOR LBN PAN PRT SVK SVN SGP SWE BEL AUT CAN CHE MDV VEN SVKSVN LCA CHL ISR FIN DNK GAB KNA HRV PRT SWE SAUAUS MYS CAN GRC CYP NZL GBR FRA DNK ISL IRL PAN VNM HKGNLD HKG SGP NOR BEL CHE BRN Log of GDP per capita Data (c) Extensive margin of exports (d) Extensive margin of imports Export Concentration Intensive Theil index HS 6 digit Mean of the index from Simulation ISR CRI SAU SEN KEN PAN BHR IRN NGA GAB GHACMR SLVHND TJK COL MUS CYP QATKWT AUS GEO LBN JOR ZAF TTO BRATUR ARG GTM AGO JAMPAN NER ARM KAZ LVAEST NZL PER TZA SYR CODOMN CHL SGP GRC LKA DOM PER BGR ROU AZE UZB EGY ECU VNM CAF VEN SWE UKR GHA CMR KGZ PHL TTO SGP ECU ZAF MRT MYS LCA ARG BLR LTUARM FIN JAM IND CHL MEX MLT HRV HUN CRI HKG IRL IRN MLT ISR NOR CAN ISL NPL UZB RUS SVK AUS BRN UGA TZA BGD THA IND BHR GMB HND BOL JOR BRA KAZ GAB MUS KNA ESPHKG ALBMDA MARBDI GEO POL PRT MYS AUT IRL MDG PHLBFA URY AZE ALB BIH ISL BLR HUN KORKWT MDG CHN DOM ESP FRA FIN KEN LAO MDV MEX RUS SEN PAK VNMPRY SAU QAT NGA LBN LVA NZL UGA MDA JPN MAR UKRURY THA ETH SLV IDN GTMTUN COL SVK SVN CYP BEL CHE BDI COD EGY FJI KGZ BIH TUN BEN USABGR EST CZEDNK GRC GBR NOR LKA LTU POL OMN PRT SWE YEM SYR VEN TUR ROU GER AUT BFA BRN CZE HRV SVN USA TKM NLD ETH KNA ITA BEL NER MWI BGD DNK CHE MRT LCAMDV TJK PAK TKM MWITCD CPV NLD ITA CPV BOL IDN GNB RWA GBR RWA FRA PRY CHN GER AGO STPYEM STP KOR LAO JPN BEN FJI CAN GMB CAF GNB NPL TCD Log of GDP per capita Data Import Concentration Intensive Theil index HS 6 digit Mean of the index from Simulation IND MLT TJK MYS HKG LCA SGP BRN MDV OMN ISR IRL USA TCD ETH AGO AZEARMALB AUT BEL BDI BEN BFA BGR CAF CPV BGD BOL BIH BRA BLRAUS BHR CAN CHE RWA GNBGMBNGA NER GHA CMR GAB GEO EGY GTM COL CRI CHLCHN CYP GRC EST ESP GBR CZE DNK MDGMRT STP KEN HND JAM HRV KGZKNA HUN FIN COD ECUDOM IND HKG FRA GER FJI IRN LBNKAZ IDN JOR ISL ITAIRL LAOLCA LKA MDA MAR LVA LTU MLT NLD MDV MUS MEX KWT ISR KOR JPN MWI NPL SEN PAK PRY PAN PHL OMN ROU RUS POL PRT SLV SYR SAUQAT MYS SVKSVN UGA NOR PER BRN ARG NZL SGP SWE TCD TZA YEM VNM TJK TUN BEL ZAF UKRTHA TKM TTO BEN URY BGD BHRCHE COD GNB PRY TUR USA UZB VEN AGO ARM AUS BFA CMRBOL AZE CHN THA EGY ECU COL BRA GAB KNA SAU LBN HUN GRC GBR KWT KEN ALB ARG CHL AUT CAN ETH GHA HND GEO JOR IRN MEX DOM CRI FIN GER IDN GTM JAM KOR JPN CZE ESP DNK NLD QAT SEN NZL NERNGA PAK LKA RUS TTO NOR MDG TZA SLVYEM ZAF VEN SVK ISL MRTMARPER KAZ PRT UGA STP VNM TUNTUR SWE MUS BDI GMB LTU MWI SYRUZB UKR BGR HRV CAFNPL KGZ CPV TKM ROU URY LVA POL ESTSVN FRA RWA MDA BIHBLR CYP LAO FJI PHL PAN Log of GDP per capita Data (e) Intensive margin of exports (f) Intensive margin of imports Figure 4: Simulated (in red) and empirical observed (in blue) export and import concentration versus GDP across 130 countries. The simulation is based on estimated trade costs form bilateral trade shares including an exported fixed effect. 26

27 Number of exporters per product Data versus simulation Red Simulated Data Blue Empirical Data Number of importers per product Data versus simulation Red Simulated Data Blue Empirical Data Frequency Frequency Number of exporters per product Number of importers per product (a) Share of products per exporting country (b) Share of products per importing country Figure 5: The simulated (in red) and empirical observed (in blue) share of the number of products traded against the number of trading countries. Export Elasticity Technology Partial versus General Equilibrium CHN USA ALB AGO GER ITA FRA ESP BRA JPN ARG IND IDN AUT BHR AUS BDI CHN CAN BEN CHE BEL CHLBGR BLR CMR GBR ISR HND FINIRL CZE DNK CRI HUN HKG COL GRCIRN EGY GTMEST BGD GHA HRV KAZ KOR KWT KEN LTU LVA MLT THA MDA MYSWE SGP TUR MEX NLD USA NOR PAKPAN PHL NGA NZL MAR JOR ECU MUS GEO PRY RUSAU POLPRT UKR ROM VNM ZAFSVKSVN VEN PER URY BFA AZE BOL ARM LAO QAT SYR TUN LKALBN TZA SLV TTO OMN ZAR SENDOM CPV CAF CYP UGA MDG ETH GER GAB YEMWI BIH NPL FRA NER ESP ISL BDI ITA MYS AGO THACAN SGP GBR CHE BEL ALB BRA JPN ARG IND AUT DNK HKG LAO IDN SWE KOR NLD IRL PAN LTU MEX EST CZE BIH FIN GRCBGR GAB GMB CRI HUN CHL EGY LVA HND MAR LKA MUS POL PRT VNM NOR RUSAU SVK PAK NGA SVN TKM KGZ TUR URY TUN HRV LBN CYPPRY BEN AUS BLR BGD GEO MLT UKR ROM TZA CAF ISR KEN NZL COL BHR GHA KAZ OMN JAM MDA IRN PHL QAT RWA SYRECU BOL GTM JOR SLV TTO YEM ZAF TCD KNA BRN KWT VEN SEN UGAAZEUZB CMR BFA MDG ETHMWI MRT NPL ZAR PER ARM NER GNB DOM KGZ STP 4 5 Red Partial Eq. CPV Blue General Eq log(share of products exported) Export Elasticity Trade Costs Partial versus General Equilibrium CHN USA ISL BDI GER ITA BRA ARG ARMALB AGO FRA ESP THA JPN IND IDN AUS BHR CHN CANFIN AUT CRI COL CHL BEL BGR BLR HND CMR BGD AZEBFA BIH CHE CZE BOL BEN CPV DNKEGY EST ECU GAB GBR KOR ISR IRL HUN GRCIRN KWT KEN LTU MUS LAO MYSWE SGP TUR RUS MEX SAU NGA GTM GHA KAZ GEO HKG LKA HRV JOR JAM MDA LVA MAR MWI NLD POLPRT NOR PAK PHL PAN NZL PER SLV UKR ROM ZAF VEN PRY RWA SVKSVN QAT SYR TUN TZAUGA TTO YEM VNM URY CYP USA DOM ETH GMB SENLBNOMN MLT UZB MDG ZAR MRT NPL CAF NER BRN GER KGZ TCD KNA FRA ESP CPV ITA MYS CAN STP THASGP GBR JPN NLD CHE BEL HKG BRA IND ARG IDN SWE KOR AUT IRL DNK VNM EST LKALBN KNA MEX NOR PAN LAO ISL MRT POLPRT FIN CZE MLT RUS GRC SVK SVN TURSAU PAKURY ALB HUN BGRLTU CYP MUS TKM NGA AGOGMB CRIEGY MAR GHA MDA STP UKR SYR HRV OMN PRY ETH AUS CHL LVA GEO ISR ROM QAT TUN NZL GAB BLR BGD BHR BRN COL GTM HND JOR TZA PHL JAM KEN NPL ZAF SEN ECU KWT SLV AZEMDGBEN IRN KAZ TTO BOL ARM VEN UGA CMR BIH GNB PER UZB DOM NER BFA YEMWI TCD BDI RWA TJK LCA 10 CAF MDV Red Partial Eq. Blue General Eq log(share of products exported) ZAR KGZ (a) Reducing technology (b) Reducing trade costs Figure 6: The simulated export concentration elasticity in the partial (in red) and the general (in blue) equilibrium. 27

28 9 Tables Table 6: Country-Specific Technology and Trade Costs estimates Country Exporter FE Standard error Precent cost Si Standard error (λ US /λ i ) θ USA AGO ALB ARG ARM AUS AUT AZE BDI BEL BEN BFA BGD BGR BHR BIH BLR BOL BRA BRN CAF CAN CHE CHL CHN CMR COL CPV CRI CYP CZE DNK DOM ECU EGY ESP EST ETH FIN FJI FRA GAB

29 Table 7: Country-Specific Technology and Trade Costs estimates - cont. Country Exporter FE Standard error Precent cost Si Standard error (λ US /λ i ) θ GBR GEO GER GHA GMB GNB GRC GTM HKG HND HRV HUN IDN IND IRL IRN ISL ISR ITA JAM JOR JPN KAZ KEN KGZ KNA KOR KWT LAO LBN LCA LKA LTU LVA MAR MDA MDG MDV MEX MLT MRT MUS MWI MYS NER NGA NLD NOR NPL NZL

30 Table 8: Country-Specific Technology and Trade Costs estimates - cont. Country Exporter FE Standard error Precent cost Si Standard error (λ US /λ i ) θ OMN PAK PAN PER PHL POL PRT PRY QAT ROM RUS RWA SAU SEN SGP SLV STP SVK SVN SWE SYR TCD THA TJK TKM TTO TUN TUR TZA UGA UKR URY UZB VEN VNM YEM ZAF ZAR

31 References ALVAREZ, F. AND R. J. LUCAS (2007): General equilibrium analysis of the Eaton Kortum model of international trade, Journal of Monetary Economics, 54(6), ANDERSON, J. E. AND E. VAN WINCOOP (2003): Gravity with gravitas: a solution to the border puzzle, American Economic Review, 93, (2004): Trade costs, Journal of Economic Literature, 42, BAS, M., T. MAYER, AND M. THOENIG (2015): From Micro to Macro: Demand, Supply, and Heterogeneity in the Trade Elasticity,. BERNARD, A. B., J. EATON, J. B. JENSON, AND S. KORTUM (2003): Plants and productivity in international trade, American Economic Review, 93, BRENTON, P. AND R. NEWFARMER (2007): Watching more than the Discovery Channel: export cycles and diversification in development, Policy Research Working Paper, CADOT, O., C. CARRÈÈRE, AND V. STRAUSS-KAHN (2011): Export Diversification: What s behind the Hump? The Review of Economics and Statistics, 93, COSTINOT, A., D. DONALDSON, AND I. KOMUNJER (2012): What goods do countries trade? A quantitative exploration of Ricardo s ideas, Review of Economic Studies, 79, COWELL, F. A. (2009): Measuring Inequality, vol. LSE Perspectives of Economic Analysis, Oxford University Press. DE FERRANTI, D., G. E. PERRY, D. LEDERMAN, AND W. E. MALONEY (2002): From natural resources to the knowledge economy: trade and job quality, Washington, DC: World Bank. DI GIOVANNI, J. AND A. A. LEVCHENKO (2012): Country Size, International Trade and Aggregate Fluctuations in Granular Economies, Journal of Political Economy, 120, EATON, J. AND S. KORTUM (2002): Technology, Geography, and Trade, Econometrica, 70(5), EATON, J., S. KORTUM, AND F. KRAMARZ (2011): An anatomy of international trade: Evidence from French firms, Econometrica, 79, FEENSTRA, R. C., R. E. LIPSEY, AND H. P. BOWEN (1997): World Trade Flows, , with Production and Tariff Data, National Bureau of Economic Research Working Paper, GAULIER, G. AND S. ZIGNAGO (2009): BACI: International trade database at the product-level,. 31

32 GRUBEL, H. G. AND P. J. LLOYD (1975): Intra-industry trade: The theory and measurement of international trade in differentiated products, vol. 12, Macmillan London. HADDAD, M., J. J. LIM, C. PANCARO, AND C. SABOROWSKI (2013): Trade openness reduces growth volatility when countries are well diversified, Canadian Journal of Economics/Revue canadienne d économique, 46, HAVEMAN, J. AND D. HUMMELS (2004): Alternative hypotheses and the volume of trade: the gravity equation and the extent of specialization, Canadian Journal of Economics/Revue canadienne d économique, 37, HESTON, A., R. SUMMERS, AND B. ATEN (2009): Penn World Table Version 6.3 Center for International Comparisons of Production, Income and Prices at the University of Pennsylvania. HUMMELS, D. AND P. KLENOW (2005): The Variety and Quality of a Nations Exports, American Economic Review, 95, IMF (2014): Long-Run Growth and Macroeconomic Stability in Low-Income Countries The Role of Structural Transformation and Diversification, IMF Policy Paper. JANSEN, M. (2004): Income volatility in small and developing economies: export concentration matters, Tech. rep., WTO Discussion Paper. KOREN, M. AND S. TENREYRO (2007): Volatility and development, The Quarterly Journal of Economics, 122, (2013): Technological diversification, The American Economic Review, 103, KRUGMAN, P. (1980): Scale Economies, Product Differentiation, and the Pattern of Trade, The American Economic Review, 16, LEVCHENKO, A. A. AND J. ZHANG (2011): The evolution of comparative advantage: Measurement and welfare implications, Tech. rep., National Bureau of Economic Research. OTTAVIANO, G. I. P., T. TABUCHI, AND J. F. THISSE (2002): Agglomeration and Trade Revisited, International Economic Review, 43(2), PAPAGEORGIOU, C. AND M. N. SPATAFORA (2012): Economic diversification in LICs: Stylized facts and macroeconomic implications, 12-13, International Monetary Fund. RAMEY, G. AND V. A. RAMEY (1994): Cross-country evidence on the link between volatility and growth, American Economic Review, 85, ROMER, P. (1994): New goods, old theory, and the welfare costs of trade restrictions, Journal of Development Economics, 43,

33 SHIKHER, S. (2011): Capital, technology, and specialization in the neoclassical model, Journal of International Economics, 83, SILVA, J. AND S. TENREYRO (2006): The log of gravity, The Review of Economics and Statistics, 88, SIMONOVSKA, I. AND M. WAUGH (2011): The Elasticity of Trade: Estimates and Evidence, Tech. rep., National Bureau of Economic Research. WAUGH, M. E. (2010): International trade and income differences, American Economic Review, 100,

34 Appendix Theil index We want to show that the total Theil index T ix = 1 N N k=1 ( ) r ik rik R ln ix R ix is the sum of the extensive and the intensive component, i.e. T ix = Tix Ext intensive component: Tix Int = 1 r N ix ik ln r k N ix ix ( rik r ix ) + Tix Int. Let s start with the Note that in the intensive component, mean export revenues are conditional on product k being exported and defined as r ix = 1 N ix r ik k N ix Next, we define the unconditional mean export revenues as R ix = 1 N k N r ik, which is defined over all products in the world, i.e. whether the product is exported or not. Substituting the unconditional for the conditional mean r ix = N N ix R ix in the intensive component, we get which we can simplify to T Int ix = 1 N k N ix Tix Int = 1 N ix N ix N and after further simplification we get T Int ix = 1 N k N ix k N ix ( ) r ik rik ln + 1 R ix R ix N ( ) r ik rik N ix ln R ix R ix N k Nix r ik 1 N ln k N r ik ( ) ( ) r ik rik N ln ln R ix R ix N ix ( ) Nix. N Now adding the extensive component to the intensive component we get the total Theil index: T Int ix + T Int ix ( ) N = ln + 1 N ix N k N ( ) ( ) r ik rik N ln ln = T ix. R ix R ix N ix 34

35 The distribution of import expenditure Note, that the set of imported goods is defined as the sum of all the fractions of goods imported from all other countries in the world. D n,k describes the fraction of goods country n imports from country k. Since a country imports a good from only 1 source country (so the sets of products imported from different countries are mutually exclusive), we can sum add together all the fractions of goods imported across all trading partners in order to obtain the import probability. Pr(imp) = I D nk k =n The corresponding distribution function for prices (p) is given by: p M n (p) = I k =n 0 N s =n (1 G ns(q))dg nk (q) k =n I D nk M n (p) = ( ) k =n I D p nk 0 ϕ n 1 θ q 1 θ 1 e ϕ nq 1 θ dq I k =n D nk Define u = s=1 N λ s( wβ s p 1 β ms κ ns ω ns ) 1 θ q 1 1 θ and du = ϕ n θ q 1 θ 1 dq = 1 u θ q dq, we get: M n (p) = ˆ p 0 (e u )du Hence, the import price distribution is independent of the source country: M n (p) = 1 e ϕ n p 1 θ = F n (p) Using the import price distribution, we can derive the distribution of import expenditure by using the following transformation. Note, that the import expenditure of country n on good x in the case of CES preferences is given by: ( ( ) ) 1 η q n (x)p n (x) = min p i(x) p η mnq n = i =n ( B wβ i p1 β mi i =n min κ ni ω ni x θ i ) ) 1 η p η mnq n (11) and the probability that it will import at price p is given by M n (p). Hence, we can write the distribution function for import expenditure at price p as E n (p) = 1 e ϕ nk 1 (1 η) n (p) 1 θ(1 η) 35

36 where k n = ( p η mnq n ) is a constant. The corresponding Fréchet pdf is e(p) = ( ) 1 1 p θ(1 η) ϕ n k n θ(1 η) ( 1 p ) e ϕ n( p kn ) 1 θ(1 η) with location parameter s n = k n 1 ϕ θ(1 η) n and shape parameter α = 1 θ(1 η).17 Given that the price distribution is independent of the source country and follows a Fréchet, we can calculate the corresponding concentration indexes analytically. The intensive margin Theil index for country n can be written approximately in terms of the continuous probability density function: T W n = [ 1 N a ( ) ( )] ˆ ( ) ( ) Rk Rk ln ln f (R)dR k G a R a R a 0 R Ra R Ra Plugging the location parameter s n = k 1 n ϕ θ(1 η) n and shape parameter α = 1 θ(1 η) density of the Fréchet distribution, we get: T W n = ˆ 0 ( ) ( ) α ln R Ra R Ra s n ( R s n ) α 1 e (R/s n) α dr into the where R a is the mean import expenditure. Solving the integral, gives: T W n = (ˆ 1 ( Γ(1 1 α ) ln u ( 1/α)) ) u ( 1/α) e u du ln (Γ(1 1α ) ) 0 (12) The Theil index for the intensive margin of imports does not depend on the scale parameter s n and is thus identical across countries. The index is completely determined by the shape parameter α = 1/(θ(η 1)) and depends only on the elasticity of substitution η and the degree of comparative advantage θ. The integral in equation 12 cannot be solved analytically. To compute the exact Theil index implied by the shape parameter α, I approximate the integral numerically via the Gauss Laguerre procedure. Trade data To build my empirical evidence, I use the BACI data set based on the Comtrade data set collected by the United Nations. I choose the digit HS product classification scheme as the preferred level of disaggregation. I assume that the tradable goods sector corresponds to the manufacturing sector. Using a correspondance table provided by Feenstra et al. (1997), I identify 4,529 tradable manufacturing products. I construct trade shares D following Bernard et al. (2003) and Waugh (2010) in the following manner: 17 The generic form of the Fréchet probability density function is: f (x) = α ( xs ) 1 α s e (x/s) α. 36

37 Imports i,j D i,j = Gross Mfg. Production i - Exports i + Imports i The numerator consists in the aggregate value of manufactured goods that country i imports from country j. These data are obtained directly from BACI. The denominator is gross manufacturing production minus total manufactured exports plus manufactured imports (against all countries in the sample), see Eaton and Kortum (2002). Basically, this simply computes the share of expenditure by dividing the value of inputs consumed by country i and imported from country j divided by the total value of inputs in country i. Gross manufacturing data are from either UNIDO (2012) or imputed from value added data obtained from the UN National accounts. Alternative classification schemes (5 digit SITC and 6 digit NAICS) In this paper I analyzed total net trade flows for the 6-digit HS industry classification. This section shows that the results obtained in the main part of the paper are influenced by the industry classification scheme and in fact apply in a more general sense. As a robustness check, I use the 5-digit SITC and 6-digit NAICS classification schemes. The total number of SITC products is 2,442 and the total number of NAICS products is 460. Table 9: Mean concentration indexes for net trade flows based on 6-digit NAICS for 130 countries over the period Concentration Gini Theil Exports Theil Imports Exports Imports Extensive Intensive Extensive Intensive Total Margin Margin Margin Margin Level Share 53% 47% 29% 71% Correlation with HS Total Regression log(gdp per capita) *** ** *** ** *** *** *** [ ] [ ] [0.0334] [0.0341] [0.0557] [ ] [0.0134] [0.0171] log(population) *** *** *** *** *** * [ ] [ ] [0.0212] [0.0264] [0.0380] [0.0107] [0.0180] [0.0216] Observations R-squared

38 Table 10: Mean concentration indexes for net trade flows based on 5-digit SITC for 130 countries over the period Concentration Gini Theil Exports Theil Imports Exports Imports Extensive Intensive Extensive Intensive Total Margin Margin Margin Margin Level Share 51% 49% 27% 73% Correlation with HS Total Regression log(gdp per capita) *** *** *** ** *** * *** *** [ ] [ ] [0.0378] [0.0366] [0.0567] [0.0172] [0.0159] [0.0218] log(population) *** ** *** *** *** *** [ ] [ ] [0.0231] [0.0247] [0.0375] [0.0149] [0.0170] [0.0211] Observations R-squared Table 10 shows the descriptive statistics for the SITC and the NAICS samples. The qualitative estimates for SITC and NAICS classifications are very similar to those of the HS classification. Exports are more concentrated than imports (Fact 1). Concentration is driven by the extensive margin for exports and the intensive margin for imports (Fact 2). Also, in terms of cross-country evidence, larger countries import and export more goods. Strikingly, the L pattern of the extensive margin also appears when the SITC and NAICS classifications are used. Poisson parameter approach The data contain intra-industry trade whereas the model is a pure Ricardian model. In this section I outline an alternative approach that converts the measurement of product units in the model to the product units in the data. Suppose that the true level of disaggregation of Ricardian products, as defined in the Eaton and Kortum (2002) (EK) model, is unobserved and the classification scheme measures only an aggregate of those Ricardian products. For example, when products, in the sense of the EK model, arrive at the border, customs agents aggregate those products into an industry according to the HS classification standard. The number of EK products that customs agents assign to an HS industry classification is unobserved to the researcher. Given this interpretation, I model the classification process as a randomization device following a Poisson process with parameter µ. The parameter µ informs on how many EK Ricardian products, on average, make up one HS code (the observed product unit in the data). To estimate the Poisson parameter, I proceed as follows. According to the law of large numbers, 38

39 the probability of importing a particular EK product equals the share of the number of EK products imported with respect to the total number of EK products. In the model, the probability that an EK product is imported equals P(imp EKprod ) = 1 D ii, where D ii is the probability of not importing an EK product. The probability of not importing any EK product within an HS code is D µ ii, where µ is the average number of products that comprise an HS code. As a result, we get the probability of importing an HS code (product unit in data), which corresponds to one minus the probability of not importing any EK products in that industry, P(imp HScode ) = 1 D µ ii. Since the probability of importing a product equals the share of products imported, N M /N, we can use the definition of the Theil index on the extensive margin, T bm i = ln(n M /N) = ln ( 1 D µ ii ) to obtain µ: ( ( ln 1 exp( T bm µ i = i ) ) ) ln(d ii ) We compute the Poisson parameter for each country and take the average value as our estimate of ˆµ = 1/I i=1 I µ i. The results imply that, on average, ˆµ = 0.94 EK products correspond to an HS code. code. Empirical evidence and simulation results In my simulation the total number of intermediate goods (N) is the product of the 4,529 industries in the data times 0.94, the average number of products in an industry, N = 4, 258. One advantage of the this approach is that we can make use of the full data sample and do not lose 35 percent of trade flows when converting the data into net trade flows. Next, I present the empirical results for the full sample together with the corresponding simulation results that replicate the 39

40 data. Table 11: Summary statistics of the average concentration indexes for 130 countries over the period Concentration Gini Theil Exports Theil Imports Exports Imports Extensive Intensive Extensive Intensive Total Margin Margin Margin Margin Level Share of total 35% 65% 28% 72% Correlation with net Total Regression log(gdp per capita) *** *** *** *** *** *** *** [ ] [ ] [0.0450] [0.0516] [0.0617] [0.0348] [0.0211] [0.0288] log(population) *** *** *** *** *** *** [ ] [ ] [0.0279] [0.0357] [0.0368] [0.0265] [0.0203] [0.0241] Observations R-squared Table 11 shows that, in general, the pattern of export and import concentration in the full sample is similar to the pattern in the net trade flow sample. Exports are more concentrated than imports for both, the Gini coefficient as well as the Theil coefficient. Regarding the quantitative differences, we observe that the overall level of concentration decreases with respect to the net trade flow sample for both exports and imports. The decomposition reveals that the effects differ across the margins. In the case of the extensive margin, concentration decreases with respect to the net trade flow sample whereas intensive margin concentration increases, thus reversing the relative importance of each margin in terms of overall export concentration. Intra-industry trade increases the number of products traded and the sales value of those products. As a result, we observe a lower (higher) concentration index on the extensive (intensive) margins. The overall concentration index is primarily driven by the intensive margin with a share of 59% for exports and 66% for imports (see Table 11). Using data on gross trade flows, I re-estimate trade cost and technology parameters. I then simulate the model, calculate the resulting concentration indexes using the Poisson measurement device and compare the simulated results with the data. Table 12 shows the results. In line with Fact 1, export concentration is higher than import concentration on all margins. With respect to Fact 2, similar to the net trade case, the extensive margin dominates overall concentration for exports and the intensive margin dominates for imports. The concentration levels obtained for imports are close to the one in the data. 40

41 Table 12: Simulated concentration level with Poisson parameter µ = 0.94 and exporter fixed effect Gini Theil Exports (X) Theil Imports (M) Extensive Intensive Extensive Intensive Exports Imports Total Model Margin Margin Margin Margin Total Simulation % 40% 39% 61% Data (gross trade flows) % 59% 34% 66% Data (net trade flows)