69764 NBER WORKING PAPER SERIES EXPORTS, EXPORT DESTINATIONS, AND SKILLS Irene Brambilla Daniel Lederman Guido Porto Working Paper 15995 http://www.nber.org/papers/w15995 NATIONAL BUREAU OF ECONOMIC RESEARCH 1050 Massachusetts Avenue Cambridge, MA 02138 May 2010 We are especially grateful to Agustin Lodola for his help with the data. We thank A. Costinot, C.T. Hsieh, E. Verhoogen, and J. Vogel for helpful suggestions and seminar participants at the NBER, Columbia University, Universidad Nacional de La Plata (UNLP), Universidad de San Andres (UdeSA), Universidad Torcuato Di Tella (UTDT), Latin American Economic Association (LACEA), the Trade Integration and Growth Network (TIGN) of LACEA, and the InterAmerican Development Bank. The views expressed herein are those of the authors and do not necessarily reflect the views of the National Bureau of Economic Research. © 2010 by Irene Brambilla, Daniel Lederman, and Guido Porto. All rights reserved. Short sections of text, not to exceed two paragraphs, may be quoted without explicit permission provided that full credit, including © notice, is given to the source. Exports, Export Destinations, and Skills Irene Brambilla, Daniel Lederman, and Guido Porto NBER Working Paper No. 15995 May 2010 JEL No. F13,F14 ABSTRACT This paper explores the links between exports, export destinations and skill utilization by firms. We identify two mechanisms behind these links, which we integrate into a unified theory of export destinations and skills. First, exporting to high-income countries requires quality upgrades that are skill-intensive (Verhoogen, 2008). Second, exporting in general, and exporting to high-income destinations in particular, requires services like distribution, transportation, and advertising, activities that are also intensive in skilled labor (Matsuyama, 2007). Both theories suggest a skill-bias in export destinations: firms that export to high-income destinations hire more skills and pay higher wages than firms that export to middle-income countries or that sell domestically. We test the theory using a panel of manufacturing Argentine firms. The data cover the period 1998-2000 and thus span the Brazilian currency devaluation of 1999. We use the exogenous changes in exports and export destinations brought about by this devaluation in a major export partner to identify the causal effect of exporting and of exporting to high-income countries on skill utilization. We find that Argentine firms exporting to high-income countries hired a higher proportion of skilled workers and paid higher average wages than other exporters (to non high-income countries) and domestic firms. Instead, we cannot identify any causal effect of exporting per se on either skill utilization or average wages. Irene Brambilla Guido Porto Calle 6 entre 47 y 48 Calle 6 entre 47 y 48 Departamento de Economía Departamento de Economía Universidad Nacional de La Plata Universidad Nacional de La Plata Argentina Argentina and Universidad de San Andrés (UdeSA) guido.porto@depeco.econo.unlp.edu.ar and also NBER irene.brambilla@econo.unlp.edu.ar Daniel Lederman The World Bank 1818 H Street NW MSN I-8-801 Washington, DC 20433 DLederman@worldbank.org 1 Introduction This paper studies whether that act of exporting affects �rm behavior and why this behavior is affected. Bernard and Jensen (1995, 1999) pioneered empirical work to characterize exporting �rms and found that exporters outperform non-exporters: they pay higher wages, they are more productive, they are larger and so on. These �ndings were later con�rmed by various authors.1 In many cases, these correlations do not imply causality (and thus cannot be used to predict changes in �rm behavior); in those cases where causality is established, the evidence tends to support a story of selection. This suggests that while exporting allows �rms to take advantage of their inherent good attributes, such as productivity, the act of exporting itself does not necessarily affect �rm behavior. In this paper, we instead identify one instance where exporting has behavioral responses in terms of skill utilization. Our work is thus in line with Verhoogen (2008), who shows that exporting induces �rms to upgrade quality, and Bustos (2009), who �nds that exporters adopt newer and better technologies than non-exporters. Concretely, we elaborate upon a theory of how exports and export destinations affect the utilization of skilled labor, and we document those features using a panel of manufacturing �rms in Argentina. Traditional theories of international trade often take a relatively simple view of the production process in which the production of “goods� is carried out by combining factors (labor, capital) and a technology. Recent trade models, including Feenstra and Hanson (1996), Matsuyama (2007), Verhoogen (2008), and Grossman and Rossi-Hansberg (2008, 2009), internalize some of the complexities of modern production processes by assuming that the production of goods comprises the combination of activities like various manufacturing tasks, marketing, distribution, foreign trade activities and exporting services. These tasks differ in their skill intensity so that the act of “exporting� becomes a skilled intensive activity, even when the act of “manufacturing� is unskilled-intensive. Moreover, we argue that the act of “exporting to high-income destinations� may require technologies and tasks that are yet more skill-intensive. In consequence, economies that trade with high-income countries will utilize relatively higher levels of skills than economies that are either closed or specialized in trade with middle- or low-income countries. There are various reasons why the act of exporting by developing countries may demand skills, 1 See Bernard and Jensen (2004), Bernard, Jensen, Redding and Schott (2007), De Loecker (2007), Schank, Schnabel and Wagner (2007), Clerides, Lach, and Tybout (1998), Pavcnik (2002), Park, Yang, Shi, and Jiang (2008), and many others. Bernard, Jensen, Redding and Schott (2007) also �nd that importers share many of the good attributes of exporters. 2 even when the production process is relatively intensive in the use of unskilled labor. A leading recent theory is provided by Verhoogen (2008), who developed a model where exports require quality upgrading—an activity that demands skilled labor. This idea can be extended to accommodate models where exporting requires “associated services� such as labeling or customer support and where the provision of these services is a skilled-intensive activity. Note that exporting per se does not necessarily require quality upgrading. A country like Argentina—the target country in the empirical work—may need to produce the same level of quality to sell internally than to sell to neighboring markets like Brazil. In contrast, exporting to high-income countries (e.g., the U.S.) does require higher quality and more skills. Our claim is that “where you export� matters.2 Matsuyama (2007) advances another reason why export destinations may require varying levels of skills. He developed a model of “skilled-bias globalization� in which international trade activities use resources and are relatively skilled-intensive. These activities include international marketing and commercialization, transportation and distribution, and perhaps also advertising (as in Arkolakis, 2009). These activities require expertise in international business, languages, foreign technologies, and social idiosyncrasies of foreign markets. In Matsuyama’s model, the technologies to supply goods depend on whether �rms sell domestically or abroad. In our setting, the technology to supply goods may also depend on the destination of exports. Once again for a country like Argentina, the activities needed to access high-income countries may require more skills than those activities needed to access neighboring markets. We test our hypothesis using a panel of Argentine manufacturing �rms, the Encuesta Nacional Industrial, ENI (National Industrial Survey). The surveys include information on sales, wage bill, employment of production and non-production workers, type of ownership, plant age, and other general characteristics of the �rms (like industry affiliation). Using con�dential information, we matched the �rms in the ENI with administrative customs data available for 1998, 1999 and 2000. The result is a combination of typical information from industrial surveys with information on export volumes and export destinations at the �rm level. In other words, we know, for each �rm in the panel, whether the �rm exported, how much it exported, and where it exported. Our 1998-2000 data span the Brazilian devaluation of 1999, which provides a nice setting for identi�cation. Brazil is a major trade partner of Argentina, and the 1999 devaluation had a large 2 Our theory is thus related to the argument in Hausmann, Hwang and Rodrik (2007), who claim that “what you export� matters. If goods are differentiated by export destination, then “what you export� and “where you export� are clearly interrelated. 3 impact on Argentine exporters. The Brazilian devaluation generated exogenous variability in the export destinations of Argentine �rms, out of Brazil and into high-income countries (and also into the domestic market). We use an instrumental variables strategy to exploit these exogenous changes to identify the role of exports and of the composition of exports in the determination of the skill composition of employment in Argentina. Moreover, since we work with a panel, we can match, for a given �rm, the behavioral changes in terms of skill utilization with the exogenous changes in its export composition induced by the Brazilian devaluation. We �nd that, for Argentine �rms, exporting to high-income countries matters, but exporting per se does not. Firms that tend to export more to high-income countries use more skills and pay higher average wages than �rms that do not export at all or export instead to middle-income countries. The reason is that the local markets in Argentina are similar to the export markets in middle-income countries and thus it is only possible to observe differences in �rm’s outcomes for �rms specializing in exporting to high-income countries. Our empirical test is related to Verhoogen’s approach, but it is more general. While Matsuyama (2007) emphasizes the technological side of the link between export destinations and skills, Verhoogen (2008) theory has a demand component (high-income countries value quality more) and a technology component (more productive �rms are more likely to produce higher quality goods). Verhoogen’s empirical test, however, explicitly exploits only the technological component of the theory by linking changes in wages for �rms with different initial levels of productivity. Our test is instead based on changes in skill utilization and exogenous changes in the destinations of exports. It thus exploits both the demand and technology features of the model and offers a more general and robust test of the theories. The remainder of the paper is organized as follows. In Section 2, we integrate the various channels linking the choice of skilled labor utilization with the act of exporting and the act of exporting to high-income destinations. In Sections 3 and 4 we lay out our empirical model, we discuss the identi�cation strategy, and we present the main results. We also discuss the main channels behind our �ndings. Section 5 concludes. 4 2 An Integrated Theory of Skills and Export Destinations The trade literature has identi�ed two broad mechanisms linking exports and skill composition at the �rm-level: quality (Verhoogen, 2008) and exporting-related activities (Matsuyama, 2007).3 In this section, we lay out a simple partial equilibrium model of trade that integrates both mechanisms and that incorporates differences among exporting markets. We �rst describe the demand side and the structure of production in a generic market. In turn, we discuss the role of different export destinations. We let products be both horizontally and vertically differentiated and allow preferences to be non-homothetic in order to capture the notion that high income countries value high quality goods more than low income countries.4 We adopt a multinomial logit utility speci�cation as in Verhoogen (2008) where consumers in high income countries have a lower marginal utility of income and thus are willing to pay a “premium� for a good of a given quality. The utility that consumer i in country of destination c derives from purchasing product j depends on a vertical differentiation parameter, denoted by θ, its price, denoted by p, and a random deviation that follows a type-I extreme value distribution, denoted by . Utility is given by (1) c c Uij = θj − αc pc + j c ij . These assumptions yield the well–known multinomial–logit aggregate demand function Mc (2) c xc (pc , θj ) = j j c exp(θj − αc pc ), j Wc where M c is the number of consumers in country c, or market size, and W c is an index that summarizes the characteristics of all available products in that market (i.e. W c = z∈Z c c exp(θz − αc pc ), where Z c is the set of available products). The parameter αc can be interpreted as the z marginal utility of income, or the inverse valuation of quality; it dictates the relative importance of θ and p in the utility function. Thus, 1/αc captures quality valuation, as in Verhoogen (2008). 3 The literature on the “quality of trade� is growing steadily. See Manasse and Turrini (2001), Hummels and Skiba (2004), Hummels and Klenow (2005), Hallak (2006), Verhoogen (2008), Hallak (2008), Hallak and Schott (2008), Hallak and Sivadasan (2009), Khandelwal (2009), Kugler and Vergoohen (2008), and Fajgelbaum, Grossman and Helpman (2009), among others. Our model combines elements from various papers in this literature. Matsuyama (2007) was the �rst to highlight the role of activities that are inherent to the act of exporting in the theory of trade and comparative advantage. 4 For simplicity, we do not consider differences in preferences among consumers in a given country. 5 The production side of our model integrates two channels linking exports and skills. One channel illustrates the skill-intensive nature of quality production and the other illustrates the skill-intensive nature of foreign trade activities. For simplicity, we assume that there are J �rms in the source country, each producing a differentiated product.5 The only input is labor and there are two types of workers, skilled and unskilled. To streamline the exposition, we assume that there is a large homogeneous goods sector that employs skilled and unskilled workers in �xed proportions. This pins down wages, and thus simpli�es the illustration of our main results. Without loss of generality we normalize the wage of unskilled workers to one and denote the wage of skilled workers by w. The delivery of a �nal good to consumers combines two sets of tasks: the manufacturing of the product and various related services such as product design, packaging, transportation and distribution, marketing research, advertising, and customer support. These two tasks are different in their skill intensity. For simplicity, we assume that manufacturing employs only unskilled workers, while the provision of services employs only skilled workers. We further distinguish between two types of services. Some services are “required services,� in the sense that they are necessary to reach consumers, such as transportation and distribution, but do not affect the value that consumers attach to a �rm’s product. We denote the level of “required services� by a parameter τ c . Differences in τ c across countries capture Matsuyama’s (2007) mechanisms. Note that in Matsuyama, τ is a variable cost, not a �xed cost. The second type of services acts as a means of vertical differentiation and thus shift the aggregate demand function for the product. The level of these “vertical differentiation services� is chosen by the �rm (via product design, packaging, advertising, customer support) and we assume they c c directly affect the parameter θj of the demand function (each �rm j chooses θj for each country of destination c). c Providing a level of services—embodied in each unit of the �nal product—of θj and τ c in c destination c requires bj (τ c + (θj )β ) units of skilled labor, with β > 1.6 The manufacturing task 5 We assume that the number of �rms is �xed, as in Chaney (2008) and Arkolakis (2008). Arkolakis, Demidova, Klenow, and Rodriguez-Clare (2008) show that this assumption yields the same results as free entry in the model of Melitz (2003). 6 To preserve Matsuyama’s Ricardian features, we assume linearity in the provision of required services, such as transportation costs, while decreasing returns to scale in the provision of vertical differentiation, such as quality of service. This latter assumption reflects the fact that shifting the demand function becomes increasingly more difficult c as quality increases. From a technical standpoint, this assumption is needed because θj is a choice variable. The skilled c c β labor requirements could also be written as bj (τ + θj ) . This alternative assumption yields very similar predictions. At the same time, we assume that the cost of providing vertical differentiation in one market is independent of the level of vertical differentiation provided by the same �rm in other markets. This assumption is in line with the constant marginal cost assumption in the manufacturing task and with independence of preferences across markets 6 requires aj units of unskilled labor per unit of the �nal product. Thus, the total variable unit cost c of providing the �nal product in country c is aj + bj (τ c + (θj )β )w. Additionally, there are �xed costs of production and of entering export markets given by F c which we describe below. While these �xed costs are common to all �rms, the labor requirements aj and bj vary at the �rm level. The relative demand for skilled workers (number of skilled workers divided by the number of unskilled c bj c workers) is Sj = aj (τ c c + (θj )β ). It is increasing in τ c and θj . For each country of destination, �rms choose pc and θj to maximize pro�ts given by πj = j c c c c [pc − aj − bj (τ c + (θj )β )w]xc (pc , θj ) − F c .7 The solutions for prices (pc ) and the level of vertical j j j c differentiation (θj ) are independent across destinations. They are given by β 1 β−1 1 (3) pc j c = aj + bj τ w + bj w + , αc bj βw αc 1 c 1 β−1 (4) θj = . αc bj βw c The vertical differentiation parameter θj is chosen so as to equalize its marginal cost with the 1 inverse of the marginal utility of income, αc . It is decreasing in the marginal utility of income (αc )—�rms choose to provide a higher level of quality when it is valued more highly —and it is c independent of τj . Price is increasing in τ c , reflecting the higher unit cost implied by a higher τ c , and decreasing in αc , reflecting both the higher unit cost implied by the higher optimal quality and the fact that �rms can extract more surplus from consumers when they are willing to pay more for their products. Given the solutions for price and vertical differentiation, we write the relative demand for skilled c workers employed in the production of goods that are shipped to country c (Sj ) as β c bj 1 β−1 (5) Sj = τc + c b βw . aj α j c Note that Sj is increasing in τ c and decreasing in αc . These features correspond to the two mechanisms outlined above. On the one hand, �rms need to provide the required services τ c because of technological requirements. On the other hand, �rms will provide services to increase (by which shifting the demand function in one country does not become increasingly more costly as demand is shifted in other markets). 7 The monopolistic competition assumption implies that �rms do not affect the index W c . 7 quality when these services are more valuable to consumers. The provision of both types of services is intensive in skilled labor. To study the role of export destinations, we assume that there are three markets: the domestic market (D), high income destinations (H), and low income destinations (L). We �rst study the mechanismsm proposed by Matsuyama (2007) and Verhoogen (2008) separately, and we later integrate them into a uni�ed theory. To explore the “required services� mechanism of Matsuyama (2007), we let (6) τ H > τ L > τ D = 0; αH = αL = αD = α. These assumptions capture the idea that exporting requires skill-intensive services, relative to production for the domestic market, and that exporting to high-income countries requires even more skill-intensive technologies. Note that, in principle, the exporting services needed to access some low-income countries can be as skill-intensive as the exporting services needed to reach high-income destinations. Exporting to the U.S. requires an English speaking manager, as does exporting to China. However, the assumption that τ H > τ L seems more plausible because doing business in high-income countries involves dealing with higher-educated consumers, more sophisticated rules of law, more stringent standards regulations and so on. Note, in addition, that our empirical analysis is a case study of Argentine �rms, where exports of manufacturing products to countries like China or India are only a small share of total exports. The comparative statics for p, θ and S are easy to derive from equations (3), (4) and (5). Since τ is higher in high income destination, �rms need to employ a higher relative skill composition in H L D order to reach these countries (Sj > Sj > Sj ). Also, since costs are higher, prices are also higher in high-income destinations (pH > pL > pD ). This is a feature of technology, not of demand. This j j j can also be seen by noting that �rms choose the same θ for both destinations, since there are no H L D differences in cross-country quality valuations (θj = θj = θj ). To explore the role of the demand (quality valuation) mechanism, as in Verhoogen (2008), we let (7) τH = τL = τD = 0; αH < αL = αD , where we assume that quality valuation in domestic markets is similar to quality valuation in 8 low-income countries. In equilibrium, �rms choose a higher level of vertical differentiation in high H L D income destinations, where quality is valued more highly (thus, θj > θj = θj ). Reaching high income countries requires a higher relative demand for skills than reaching both low-income and H L D domestic consumers so that Sj > Sj = Sj . Price is higher in high-income countries due to two reasons: higher cost (due to higher quality) and price discrimination (thus, pH > pL = pD ).8 j j j We can now combine the technology and demand mechanisms. Since both channels go in the same direction, they reinforce each other. There is skill-bias in the composition of employment for �rms that target high-income countries. Relative demand for skilled workers at the �rm level can be written as a weighted average of the relative demands for skilled workers employed in the production and delivery of goods that are shipped to each type of destination: D xD j xL L j xH H j (8) Sj = Sj + Sj + Sj . xj xj xj In this equation, x = xD + xL + xH are total �rm sales (including domestic sales and exports to different destinations). Since S H > S L > S D , it follows that the higher the share of exports that goes to high income countries, the higher the relative demand for skilled labor at the �rm level. Note also that �rms exporting to high income countries will incur a higher wage bill since skilled labor earns a higher wage than unskilled labor.9 We now turn to market entry. Firms enter a market when variable pro�ts given the optimal c decision for θj and pc are enough to cover the �xed cost F c . In our model, there are two dimensions j of �rm heterogeneity. Firms differ in a, the efficiency in the use of unskilled labor, and in b, the efficiency in the use of skilled labor. Since we assume a �xed-proportion technology, both a and b (not just their ratio) matter in sorting out which �rms have the necessary attributes to enter a market.10 Firms with lower a and b are more efficient and are thus able to enter a market. 8 Note that mark-ups are not constant in this model. However, the result of vertical differentiation could be achieved with constant mark-ups as well, with a CES utility speci�cation as in Hallak (2006) 9 There are other mechanisms that could in principle explain a positive link between exporting to high-income countries and wages in a country like Argentina. One explanation is “pro�t sharing� in a model of fair wages. Firms that export to high income countries are likely to make extra-pro�ts, which are shared with workers. See Egger and Kreickemeier (2009) and Amity and Davis (2008). It is also possible that exporting to developed countries is associated with higher wages to reduce labor turnover. This is an efficiency wages story, but it could also be interpreted as a dimension of quality because the �rm needs to reduce labor turnover to guarantee timely delivery of goods and maintain product standards. Another theory due to Yeaple (2005) argues for higher wages due to scale economies attached to exporting (to different destinations). In this model, the size of the market and the scale of the �rm determine the choice of technology and larger �rms choose more skill intensive technologies that pay higher average wages. 10 This is a useful simpli�cation that allow us to illustrate the role of multi-dimensional heterogeneity in models of trade. See Hallak and Sidivasan (2009) or Bernard, Redding and Schott (2009) for full general equilibrium models. 9 Figure 1 Entry into Export Markets Multi-Dimensional Heterogeneity b 6 Out In π c (a, b|F c , τ c , αc ) = 0 - a The zero-pro�t curve is the set of combinations of a and b the leave �rms indifferent between staying in or out of market c. More precisely, there is a trade-off between a and b, and �rms with low a can compensate for a high b (but only to a limited extent). To see this, in Figure 1, we plot a zero-pro�t curve, a level curve for the function π c = 0, where π c are indirect pro�ts in market c. The zero-pro�t curve is negatively-sloped (indicating the trade-off between a and b) and convex (indicating the increasing difficulties in trading-off one source of productivity with the other). Firms with parameters a and b below the zero-pro�t curve enter the market and the others stay out. The position of the zero-pro�t curve for market c depends on F c , τ c and αc . An increase in any of the three parameters implies a decrease in pro�ts and thus shifts the zero-pro�t curve downwards. The downward shift is parallel for an increase in F , while an increase in τ or α cause the zero-pro�t curve to become less steep. As these parameters differ across the three types of destinations, there are three zero-pro�t curves, one for the domestic market, one for low income export markets, and one for high income export markets. We assume that �xed costs are lowest in the domestic market and highest in the high income market, that is, F D < F L < F H . This assumption, works in the direction of positioning the zero-pro�t curve for the domestic market above the zero-pro�t curve for the low-income export market, which in turn is above the zero-pro�t curve for the high-income export market, as depicted in the left panel of Figure 2. Firms with parameters a and b below the curve π H = 0 enter the high 10 Figure 2 The Choice of Export Destinations b 6 b 6 Out Out D L,D D L,D πD = 0 πD = 0 πL = 0 H,D πH = 0 H, L, D πH = 0 H, L, D πL = 0 - - a a D=entry into the domestic market only; L,D=entry into the domestic and low-income markets; H,D=entry into the domestic and high-income markets; H,L,D=entry into all markets. income market. They also sell their product in the low income market and the domestic market. Intermediate �rms characterized by a combination of a and b between the level curves π H = 0 and π L = 0 enter the low income and domestic markets; while �rms between π L = 0 and π D = 0 serve the domestic market only. The least productive �rms stay out of all markets. Export technology differences across markets (τ H > τ L > τ D = 0) work in the same direction, and they affect the relative slopes of the curves as well, so that the zero-pro�t curve for high-income markets is the least steep (as in Figure 2). The vertical differentiation channel (αH < αL = αD ), on the other hand, pushes π H = 0 upwards so that it is theoretically possible for π H = 0 to intersect with π L = 0 (the right panel in Figure 2) or with both π L = 0 and π D = 0. That is, �rms that are sufficiently inefficient in the use of unskilled labor but relatively efficient in the use of skilled labor could �nd it pro�table to export to high-income markets but not to low-income markets (or even domestic markets).11 Thus, in our model, efficiency differences in the use of skilled labor are an important factor explaining the choice of export destinations (as are the country speci�c �xed costs of Chaney (2008) and the �rm-country-product consumer tastes of Bernard, Redding, and 11 Whether the zero-pro�t curves intersect or not depends on several parameters. Factors that work in the direction of curves crossing are large differences in α across destinations, small differences in F and τ , differences in number of consumers and �rms, and low relative wages of skilled workers. In our data from manufacturing �rms in Argentina, we �nd that, in 1998, 3 percent of exporting �rms export only to high-income destinations and 11 percent of �rms export only to upper-middle income destinations. As expected, all exporting �rms sell in the domestic market as well. 11 Schott, 2009). As we will shortly discuss, our empirical strategy exploits exogenous changes in exports and in export destinations of Argentine �rms brought about by a Brazilian devaluation that took place in 1999. In the theoretical model, the Brazilian devaluation causes a downward shift in the level curve π L = 0, without affecting π H = 0. As a result, some exporters to low income foreign markets will exit (eventually retrenching into local markets) and exporters to high income markets will sell less in low income markets. This implies that the share of exports to high income destination will increase and so will the utilization of skills, as per equation (8). Note that, in this setting, this is a within-�rm composition effect, where a weighted average (total skill utilization) changes due to changes in the participation of each destination in total sales. This corresponds to changes in the c weights with constant Sj in equation (8). More generally, in a model with increasing marginal costs or capacity constraints, for instance, the devaluation of the low income country’s currency could also boost exports into high income markets. In this case, the increase in the utilization of skills will be twofold, �rst because of export destination composition effects (changes in the weights in c (8)) and second because of a higher demand of skills (changes in Sj in (8)). 3 Empirical Analysis We now turn to our empirical analysis. We describe the data, introduce the regression model and the identi�cation strategy, and present the main �ndings on exports, export destinations and skill utilization in Argentina. 3.1 The Data We use two main sources of data in our analysis: a �rm survey and administrative customs information. The �rm survey is the “Encuesta Nacional Industrial� (ENI) or National Industrial Survey. The ENI is a panel of manufacturing plants and collects information on sales, value added, input use, employment of production workers, employment of non-production workers, total wage bill, and industry affiliation at the 3-digit ISIC level Revision 2. We have access to the module of the survey that corresponds to the province of Buenos Aires for the years 1998, 1999 and 2000.12 The second source of data for our analysis is administrative customs records. From the customs 12 Being the most highly industrialized and developed area of the country, the province of Buenos Aires accounts for 40 percent of the population and more than half of the manufacturing activity in terms of employment and output. 12 records we extracted the total value of exports by country of destination at the �rm level.13 We then matched this information to the �rm survey using tax identi�cation numbers. The result is a rare panel of employment, wages, exports, and export destinations by �rm. When matching the two data sources, one issue to consider was that the �rm survey was collected at the plant level while the customs information was recorded at the �rm level. Since all plants owned by a same �rm share the same identi�cation number, we aggregated information across plants owned by a given �rm and created a dataset at the �rm level that we then matched with the customs records. In our survey, only 14 percent of �rms own more than one plant. We later show that regression results are very similar whether we use the full sample of �rms or only one-plant �rms. Table 1 presents summary statistics from the combination of �rm and customs data for the full sample 1998-2000. In Panel A, we describe the export intensity and export destinations of Argentine �rms; in Panel B, we focus on differences in outcomes (employment, wages, and skill utilization) across those �rms. Out of 901 �rms and 2544 �rm-year observations, 68 percent of �rms exported in at least one of the three years of data, while in a given year, the average share of exporters is 59 percent (Panel A). The proportion of exporters is higher than what the literature typically �nds. This can be partly explained by the fact that our �rm survey corresponds to an export-oriented geographic area of Argentina.14 The share of exports in total sales is, in contrast, small: exports account for only 8 percent of sales across all �rms (column 1) and 13 percent among exporters (column 2). The average number of destinations (including the domestic market) is 3.3 in the sample of all �rms and 4.9 among exporters. In columns 3 and 4, we describe the characteristics of �rms that export to at least one high income destination, the “high-income exporters.� We work with two de�nitions of “high-income� destinations based on the World Bank country classi�cation. In the �rst de�nition (De�nition I), we include countries classi�ed as high-income OECD, high-income non-OECD and upper-middle income. The countries in each group are listed in the appendix. On average, each year 51 percent of �rms export to at least one high-income destination (1307 out of 2544 �rms). For the “high-income� exporters, total exports (to all destinations) account for 15 percent of total sales and the average number of destinations, including domestic sales, is 5.4. 13 We do not have access to quantities or unit values. 14 At the same time, the high share of exporting �rms is not explained by biases in �rm size, as the ENI covers �rms of all sizes (provided they are in the formal sector), based on censuses sampling weights. While tax evasion and worker informality are prevalent among manufacturing �rms in Argentina, �rm informality is not. 13 In the second de�nition (De�nition II), we classify as high-income destinations only countries in the World Bank’s high-income OECD and non-OECD groups, while upper-middle income economies are excluded. In this grouping, the number of �rms that export to high-income countries drops to 27 percent (680 out of 2544 �rms). These �rms show a higher export intensity (exports account for 22 percent of total sales) and a higher number of destinations (7.8 per �rm). In our regression analysis, we choose the �rst classi�cation as our main speci�cation to show that even under a conservative de�nition there are signi�cant differences between high- and low-income exporters. As we will show below, these differences are exacerbated when we apply the more liberal de�nition that excludes upper-middle income economies from the high-income destinations set. In Panel B of Table 1, we begin exploring the relationship between skill utilization, exports and export destinations by comparing some differences between exporters and non-exporters. From the ENI panel survey, we report in column 1 average employment, sales and measures of skill utilization. We consider three measures of skill utilization: average wage—de�ned as the total wage bill divided by total employment,—the share of non-production workers in total employment, and the share of hours worked by non-production workers in the total number of hours worked by all employees.15 In our sample, �rms employ an average of 89.7 workers and pay average annual wages of 12,154 USD. Non-production workers account for 26 percent of total employment, whether measured in number of workers or hours worked. In column 2, we report differences between exporters and non-exporters by running an OLS regression of each �rm attribute on a dummy of whether the �rm exports or not (which we build using the matched customs data), controlling for industry and year effects. Our data con�rm the stylized fact of this literature: Exporters are larger by around 122 percent (173 percent in terms of sales); they pay higher wages by about 48 percent; and they hire 5 percentage points more non-production workers than non-exporters. In column 3, we compare high-income exporters to low-income exporters. Conditional on exporting, we run OLS regressions of the various outcomes on a dummy for high-income exporters (controlling for industry and year effects). Using De�nition I, the results show that �rms that export to at least one high-income destination are 39 percent larger in terms of employment and 54 percent larger in terms of sales than exporters that only export to low income countries; they pay higher wages by about 12 percent; and they hire a larger fraction of non-production workers 15 Note that we do not have information on the educational level of the workers. In the surveys, �rms report the total wage bill, which includes the total payment to workers of all types, production and non-production. However, �rms do separately report the total number of employees in each of those two categories. The average wage is a proxy for skill utilization inasmuch as �rms with a higher skill composition pay higher wages. 14 by about 3 more percentage points. In short, our data reveal that while exporters have good attributes, high-income exporters have yet better attributes. In column 4, we compare high-income exporters in De�nition I with those in De�nition II. We keep �rms that export to high income countries according to the �rst de�nition and show the premium for �rms that specialize in exports to only high-income OECD and non-OECD destinations (compared to upper-middle income destinations). The data show that �rm differences persist among high-income exporters. For instance, De�nition II high-income exporters are larger (by 43 percent in terms of employment and by 54 percent in terms of sales) than De�nition I high-income exporters. They also pay higher average wages (by 15 percent) and seem to employ a higher share of non-production workers (by 2 percentage points). 3.2 Skills and Export Destinations: The Model The statistics reported in Table 1 uncover the basic relationship between exports, export destinations and skills. We now study this relationship in more detail with the following regression model: (9) sit = δ1 EXPit + δ2 HIit + xit β1 + φt + φi + it . The variable sit is a measure of the utilization of skills in the labor force employed by �rm i at time t. As explained above, we use the average wage and the share of non-production workers in total employment as proxies for skill utilization. The right-hand side variables of interest are EXP and HI. Let Eit be total exports of �rm i, let Yit be total sales (including domestic sales and exports) and let EHit be exports to High-income destinations.16 We de�ne EXPit = Eit /Yit as the ratio of total �rm exports to total �rm sales. This variable is a better measure of export exposure than an exporter dummy, since it captures the intensity of the exporting status. Furthermore, from a practical point of view, a continuous variable has much more variability within �rms in a 3 year panel such as ours. HIit = EHit /Eit is de�ned as the share of �rm exports to high-income destinations on total �rm exports. This variable is a measure of the composition of exports across destinations and captures the impact of exporting to 16 In most of this section, we use the conservative de�nition (De�nition I) where high income destinations are countries in the high-income OECD, high-income non-OECD and upper-middle income groups of the World Bank classi�cation. We later show robustness results dropping the group of upper-middle income countries from high-income destinations. 15 high-income developed countries once export intensity has been accounted for.17 The vector of �rm characteristics, x, includes industry dummies (when �rm �xed effects are not included), �rm size measured by the log of total sales, and differential trends across time at the �rm level. The differential trends are constructed by interacting sales in the �rst year of data (1998) with year effects (and exchange rates). The error term includes a �rm �xed effect φi , a year �xed-effect φt , and a random component it . We start by estimating (9) by OLS, pooling all years of data but including industry and year �xed effects. Results are in Table 2. In Columns 1 to 3, we work with our �rst proxy for skill utilization, average wages. In column 1, where we include EXP but exclude HI, we con�rm the cross-section result: �rms with higher ratios of exports to sales pay higher wages. In column 2, where we include HI, high-income destination exports, but we exclude EXP , we �nd a positive and signi�cant coefficient as well. In column 3, we include both the exports to sales ratio (EXP ) and the ratio of exports to high income (HI) in the same regression. Both coefficients are positive and statistically signi�cant. Overall, thus, we observe that skill utilization is positively correlated not only with export intensity but also with the destination of a �rm’s exports. This means that, conditional on the same export intensity, �rms that ship a larger share of their exports to high-income markets utilize, on average, more skills. In columns 4 to 6, we work with our second proxy for s, the share of non-production workers in total �rm employment. When we only include EXP (Column 4), we �nd a positive correlation, albeit not statistically signi�cant, between export intensity and skills. When we only include HI, we �nd a strong positive correlation between exporting to high-income countries and skill composition. Finally, when both EXP and HI are included (column 6), we �nd that while exporting to high-income countries is positively correlated with a higher utilization of skills, export intensity is not. Both the simple correlations in Table 1 and the OLS estimates in Table 2 are consistent with the claim that exporters utilize relatively more skilled labor and that high-income exporters even more than exporters. In the light of our model in Section 2, there are two ways to interpret these �ndings with different implications about the behavior of �rms that are engaged in exports (to different destinations). 17 An alternative speci�cation that would also allow us to test our claim is to include exports over sales as well as exports to high-income countries over sales as regressors. These variables are highly collinear, however, and the model in (9) is therefore preferred. 16 Our model proposed differences in the �xed costs of exporting to different destinations, multi-dimensional �rm heterogeneity (the parameters a and b), and differences in both the marginal utility of income (α) and in the costs of exporting (τ ) across destinations. Consider �rst a scenario where there are no differences in α and τ , so that we “shut down� both the quality (Verhoogen 2008) and exporting technology (Matsuyama 2007) mechanisms. We retain the differences in a and b and also in the �xed costs (with higher �xed costs in the high-income market, lower in the low-income market, and even lower domestically, F H > F L > F D ). In this scenario, since α and τ are the same across destinations, a given �rm will choose the same level of skills for each country of H L D destination that it exports to (Sit = Sit = Sit ). However, this level of skill will vary across �rms according to their efficiency in the utilization of skilled and unskilled labor (the parameters a and b). This is because more efficient �rms will choose a higher level of vertical differentiation for their products, which provides higher pro�ts. At the same time, the differences in �xed costs generate a sorting of �rms as described in Figure 2, where �rms with lower a and b will be more likely to �nd it pro�table to export to high income destinations. This scenario generates a positive association between the skill (s) and the export destination variables and it is thus consistent with the OLS results of Table 2. There is no causality, however. The correlation arises because of the selection of �rms into the different export markets. In other words, there are productivity and cost shocks that allow �rms to enter or expand their export operations and hire different skill levels at the same time. Let us now add the quality and export technology channels. We showed that in a scenario with lower marginal utility of income in high-income destinations (αH < αL = αD ) and with more costly technology to export to high-income markets (τ H > τ L > τ D = 0), �rms will use a higher share H L D of skilled workers in their exports to high-income markets (Sit > Sit > Sit ). This result is not directly testable with our data, since we do not observe skill use by country of destination. Instead, we observe the total skill intensity, which is a weighted average of the skill use in each destination as H D L given by equation (8). However, since Sit > Sit > Sit , we should observe that for a given �rm, the average skill use increases with the share of high-income exports. This scenario is also consistent with the OLS results of Table 2 but the implications are different. The correlation between skill use and high-income exports is not merely due to selection across �rms but also because of differences in �rm behavior (i.e., the utilization of different skills in serving different markets) in response to differences in α and τ . 17 For our analysis, we want to be able to discern between the two scenarios because they carry implications regarding the relevance of the mechanisms outlined in our theory. Clearly, it matters if the correlations between the variables are only due to selection of more productive �rms or also due to behavioral responses. OLS cannot distinguish between the two scenarios.18 To differentiate them, we develop an empirical strategy where we look at changes within �rms and use instruments that explain exogenous shifts in the export share of high-income destinations. Our strategy provides an answer to the following question: for a given �rm, with given a and b (possibly changing over H L D time), and thus given Sit , Sit and Sit , does the average skill use go up if there is an exogenous increase in the share of high income destinations? 3.3 Identi�cation Strategy There are two endogenous variables in our model: exporting to high-income destinations HI (the share of exports to high income countries on total �rm exports), and export intensity EXP (the share of exports in sales). To achieve identi�cation, we exploit the panel nature of our data and we use instrumental variables. The panel allows us to track �rms across time. To build the instruments, we exploit the exogenous variation in HI and EXP created by the Brazilian devaluation of 1999, which induced Argentine �rms to cut sales in Brazil and to expand sales both domestically and in high-income countries.19 In the end, our strategy boils down to tracking changes in skill utilization for a given �rm, given its exogenous responses in exports and export destinations following the exogenous Brazilian devaluation. Argentina and Brazil are major trade partners and thus the Brazilian devaluation had an impact on Argentine exports that is large enough to achieve identi�cation. Argentine export statistics by country of destination provide prima-facie evidence in support our strategy. In Table 3, we report that in the pre-devaluation year of 1998, Argentine exports were destined mostly to Brazil (36 18 There are other sources of unobserved heterogeneity which are beyond the scope of the model but that could be affecting the correlation in skill utilization and high income exports in the data. One important unobserved factor is imports. Bernard, Jensen, Redding and Schott (2007) show that exporters are also importers and that they share many good attributes. Another scenario, where the OLS estimates are attenuated, is when �rms are subject to policy shocks or domestic regulation shocks. For instance, �rms that are more likely to be “captured� by unions could be less likely to export while at the same time be required to pay higher wages, on average (see Galiani and Porto, 2009). Finally, if �rms that are more oriented to the domestic market must incur costs related to non-tradable operations (like distribution), then high local costs may cause some �rms to face a higher wage bill and become even less competitive in international markets. 19 There is a growing literature that looks at changes in major trade partners as a source of identi�cation. Exchange rates of trade partners were used for instance by Revenga (1992) and Park et al. (2008). Changes in Brazilian tariffs due to Mercosur were used to identify impacts on Argentine �rms in Bustos (2009). Verhoogen (2008) uses the own devaluation of Mexico to link exports to wage inequality. 18 percent), Europe (13 percent), the United States (10 percent) and neighbors like Chile (6 percent), Uruguay (4 percent) and Paraguay (3 percent). In 1999, when the crisis hit and Brazil devaluated, the share of exports to Brazil dropped to 28 percent. These shares partially recovered in 2000, reaching 31 percent. Consistent with our argument above, alternative markets for Argentine exports were found in the U.S. (with shares increasing to 13 percent in 1999 and 15 percent in 2000) and Europe (with shares increasing to 15 percent in 1999 and 14 percent in 2000). At the bottom of Table 3, we observe that the share of exports destined to High Income countries increased from 43 percent in 1998 to 50 percent in 1999 and 51 percent in 2000 (using De�nition I) and from 28 percent to 34 percent (using De�nition II). In Table 3, we also report changes in export volumes. As expected, exports to Brazil declined, because Brazilian domestic products became relatively more inexpensive. At the bottom of the table, we show that exports to high income destination actually increased signi�cantly, from 6.5 billion dollars in 1998 to 7.2 billion dollars in 2000. This expansion is largely accounted for by an increase in exports to the U.S., from 1.6 to 2.2 billion dollars. While it is interesting to con�rm that exports to high income countries increased, this result is not necessary for identi�cation. As we showed in the theoretical model (see the last paragraph of section 2), for our purposes it suffices to observe within-�rm changes in the share of high income countries in total exports after the Brazilian devaluation (independently of the changes in export volumes). Clearly, the fact that exports to high income countries actually increase in the data only reinforces the mechanisms that we exploit in our empirical analysis.20 3.4 IV Results HI We build separate instruments for HI and EXP . Our instrument for HI is de�ned as Iit = P ostt ∗ λBRA , that is, it is the interaction of a P ost devaluation variable with the pre-devaluation share i98 of the �rm’s exports that were destined to Brazil, λBRA . Since the 1998 shares λBRA precede the i98 i98 devaluation, they measure exogenous exposure to the devaluation. The rationale for this instrument is that, as documented in Table 3, following the devaluation, �rms that were mostly exposed to the Brazilian devaluation adjusted by moving away from this market and by exploring new markets in high income countries. We expect a positive correlation between the “scope to divert exports� and 20 Incidentally, note our model assumes constant returns to scale and independent demands across destinations, and thus �rms solve their optimization problems independently for each market. Since we observe changes in export volumes across destinations, we infer that there are factors linking �rm decisions across markets, such as increasing marginal costs or capacity constraints. This observation does not affect the validity of our empirical strategy. 19 exports to high-income countries. We adopt two speci�cations for P ost. In the relatively more non-parametric model we interact the level of exposure to Brazil before the devaluation, λBRA , with 1999 and 2000 year dummy i98 variables, so that the instrumental variables are HI (10) Iit 1 = φt ∗ λBRA , i98 where φt denotes the 1999 and 2000 year dummies. This speci�cation allows the impacts of the devaluation to vary from one year to the other as �rms adjust to the exchange rate shock. In the second speci�cation, we interact λBRA with the exchange rate of the Brazilian to the Argentine i98 currency in 1999 and 2000, erateBRA : t HI (11) Iit 2 = erateBRA ∗ λBRA . t i98 To deal with the endogeneity of the ratio of exports over sales (EXP), we construct a measure of the average exchange rate faced by a given �rm in international markets: EXP (12) Iit = c eratec ∗ ψi98 , t c c where ψi98 is the share of exports of �rm i to country c on total sales in 1998 (which is predetermined) and eratec is the exchange rate of country c (to the Argentine Peso) at time t. t Instruments such as (12) have been used before by, for example, Revenga (1992) and Park et al. (2008). Given the shares of export sales to market c in 1998, a higher exchange rate would induce �rm i to export more to this market, thus increasing EXPit . In consequence, we expect EXP to be positively correlated with I EXP in the �rst stage regressions.21 In the Appendix, we explore robustness results using a different instrument for EXP . Good instruments have to be exogenous, help to explain the endogenous variables, and satisfy the exclusion restrictions. Our instruments satisfy all these conditions. First, the Brazilian devaluation generated exogenous variation in export intensity (EXP ) and in export destinations (HI). These changes in exports are exogenous to the pre-devaluation shares of exports to Brazil. 21 c Note that while λBRA in equations (10) and (11) is the share of exports to Brazil in total exports, ψ98 in equation 98 (12) is the share of exports to country c on total sales. As a consequence, the instruments are not perfectly correlated and in fact convey different useful information for identi�cation purposes. 20 Likewise, our second instrument is based on arguably exogenous changes in the exchange rates of all trading partners and on each �rm exposure to those changes given their pre-shock export/sales shares. In other words, while the pre-shock shares are a choice variable of the �rm, once they are predetermined, the differential change in exports due to the devaluation of Brazil (or other countries) is exogenous. Second, as argued above, the instruments are correlated with the level of exports and its composition/destination. We test this below with results from the �rst-stage regressions. Finally, the exclusion restrictions require that our instruments do not have an effect on skill utilization beyond the indirect effect via exports and export destinations. One potential violation of the exclusion restrictions is given by the macroeconomic effects of the devaluation of a major trading partner. To account for this, we control for any direct effect of the devaluation with year effects.22 Finally, note that our strategy can fail if there is serial correlation in the errors but we cannot do much about this with our short 3-year panel (Arellano, 2003). We now turn to our IV estimates. We begin with Table 4, which shows results from the baseline model where our measure of skills is the average wage bill, the instrument for HI is the interaction between initial shares in 1998 with year dummies (equation (10)), and the instrument for EXP is the average weighted exchange rate (equation (12)). Panel A reports the IV estimates of δ1 and δ2 in (9); Panels B1 and B2 document the �rst stage regressions for HI and EXP , respectively. In column 1, we run a simple model which omits the year effects, which are added in column 2. We �nd that exporting per se does not raise skill utilization, but exporting to high income countries does. Results are robust to the inclusion of the year effects in column 2. The purpose of this exercise is to informally test for the exclusion restriction. The year effects account for the macroeconomic impacts of the devaluation (and other time effects that affected all �rms in the same fashion). If the devaluation is having a direct effect on the changes in export behavior of Argentine �rms and this response depends on the initial shares in 1998, then the coefficients attached to the instruments in the �rst stage should change when we move from column 1 to column 2 (Panels B1 and B2). However, the results in both speci�cations are very similar and this lends support to our contention that the exclusion restriction is not violated by our estimation strategy. In all remaining speci�cations, we keep the year effects as they are crucial controls in our regressions.23 22 It is important to emphasize that we are not interested in the impacts of the devaluation nor do we claim that we can identify those impacts. Instead, the argument is that identi�cation relies on exogenous changes in export destinations (towards high-income countries and also towards domestic markets) caused by the Brazilian devaluation. 23 Note that we cannot rule out a direct effect of the devaluation on export behavior. This is in fact captured by the year dummies. What matters for our strategy to work is that this is not confounding the changes in exports 21 We use results in Panels B1 and B2 to further diagnose the instruments. Overall, the instruments work very well: They have substantial explanatory power and are statistically signi�cant. We also report very low p-values associated with the F -statistic of joint signi�cance of the instruments, which consequently pass the Staiger and Stock (1997) test for weak instrumentation. In columns 3-5 of Table 4, we report robustness results where we control for differences in initial conditions in order to rule out unobserved factors that could simultaneously determine the choice of export shares to Brazil in 1998 and the subsequent response to the shock. Pre-devaluation productivity shocks or cost shocks that persist in time are leading candidates. Those unobservables could invalidate our IV strategy because they imply that a �rm’s ability to change export destinations may depend on the initial share exported to Brazil (in 1998). In other words, our assumption that even though the choice of export markets is endogenous, the change caused by the Brazilian devaluation is not, could fail if �rms more exposed to Brazil are somehow more (less) able to change destinations. While this is unlikely to be an issue in practice, we can address it directly by including controls for these unobserved pre-shock differences that may drive the potentially endogenous responses. To control for initial conditions, we interact log sales in 1998 with year dummies (column 3) and with the Brazilian exchange rate (column 4). In these speci�cations, EXP is never statistically signi�cant, while HI is always positive and highly statistically signi�cant. In addition, the coefficients in columns 3 and 4 (0.321 and 0.338) are slightly larger, but comparable, to those in the simpler models of columns 1-2. In Panels B1 and B2, we �nd that the addition of controls for the �rm’s initial conditions does not affect the statistical properties of our instruments. In column 5, we estimate the model with the log of sales as an additional control in the regression for time-varying heterogeneity such as current productivity or cost shocks. Note that we use sales as a proxy for unobserved characteristics in order to improve the estimation of δ1 and δ2 ; we consequently do not attach any causal interpretation to the coefficient of log sales. Our main results remain unchanged: EXP does not affect skill use, but HI does. The estimated coefficient of HI is 0.313, indicating that a �rm with the average shares of exports to high-income countries (30 percent) pays around 9.4 percent higher wages than �rms that do not export at all to developed countries. This is a sizeable difference, both statistically and economically. As another robustness check, we run all the previous speci�cations using the alternative conditional on the initial shares. 22 instrument for HI, equation (11) instead of (10). That is, the instrument is the interaction of the initial shares of exports to Brazil with the Brazilian exchange rate. Results are reported in Table 5. As before, Panel A shows the IV estimates of δ1 and δ2 , and Panels B1 and B2 show the �rst-stage results for HI and EXP , respectively. All our �ndings remain unchanged. The ratio of exports to sales (EXP ) is never signi�cant and the ratio of export to high income on total exports (HI) is always signi�cant. The estimated coefficients are also very similar. For instance, in our preferred speci�cation with controls for log sales (column 5), our estimate of δ2 is 0.314 (it was 0.313 in Table 4). In the �rst-stage results in Panels B1 and B2, the instruments show the expected sign, are jointly very signi�cant, and there is no risk of weak instrumentation. Before discussing our interpretation of these �ndings, we report the results from models where skill utilization is measured with the share of non-production workers on total employment. Results are in Tables 6 and 7. In Panel A, we use the ratio of the number of non-production workers to the total number of workers; In Panel B, we use the ratio of hours worked by non-production workers to the total number of hours worked by all employees. In Table 6, we report results using the instrument for HI built with year dummies (equation (10)) and in Table 7, we replace year dummies with exchange rates (equation (11)). In both cases, the instrument for EXP is given by (12). Our results suggest that, as before, the ratio of exports to sales does not seem to have an effect on the composition of skills at the �rm level, whereas exporting to high income countries does matter. In our preferred speci�cation, column 5, we �nd that a �rm exporting to high-income destination at the average share of 0.30 utilizes 3.63 percentage points more skilled workers than non-high-income exporters. We �nd the same result when we use hours of skilled labor worked rather than physical units of skilled labor (see Panel B). The reason why exporting to high-income countries (HI) is signi�cant while simply exporting (EXP ) is not can be found in Figure 2. If the domestic market in Argentina is similar to export markets in low- and middle-income economies (including any �xed costs), then the nature of domestic �rms and of low-income exporters, in terms of their attributes a and b, will be similar and, consequently, differences in skill utilization will be small. In contrast, exporting to high-income countries does need quality upgrades and required services, which implies a signi�cantly higher utilization of skills. We can learn more about the structure of manufacturing exports in Argentina by looking at the estimates of EXP . In our model, the predicted sign of the impact of export intensity on 23 a skill utilization depends on the relative values of α and τ in Brazil vis-`-vis other trading partners (since the Brazilian devaluation is basically shifting exports out of Brazil and into both high-income countries and the domestic market). If Matsuyama’s mechanism holds, then τ BRA > τ ARG (and actually τ ARG = 0). Consequently, conditional of HI, export intensity should have a positive impact on skill utilization. On the other hand, if we plausibly assume that αARG < αBRA , then (once again, conditional on the structure of exports to high-income countries), export density could negatively affect the utilization of skills (because �rms choose a higher degree of product differentiation to sell in local markets). These are conflicting forces and it is thus not surprising that our estimates of EXP are not statistically signi�cant. Note, however, that the estimates are negative and relatively large (but without enough precision) and this is (weakly) consistent with the quality mechanism. Finally, we need to discuss the issue of timing and �rm adjustments. Our data span the 1998-2000 period and hence the estimated impacts are based on changes in �rm behavior taking place in the 2-year period following the Brazilian devaluation. A concern is whether it is plausible for �rms to adjust the utilization of skills in such a short period of time. As explained above, for our test of relative skill utilization across destinations to work it is sufficient to observe compositional c changes within �rms (that is, changes in the export participation weights of the different Sj in equation (8)). Therefore, it is indeed possible to identify the mechanisms outlined in our theory in a short panel without necessarily relying on large �rm adjustments. 3.5 A Validation Exercise In this section, we perform a validation exercise where we explore results under the more stringent de�nition of “High-Income� destinations, which includes only High-Income OECD and High-Income non-OECD countries (De�nition II, see section 3 and the Appendix). Table 8 reports IV results using both log average wages (columns 1-3) and the share of non-production workers (columns 4-6) as our measures of skill utilization. To streamline the exposition, we only report IV estimates from regressions models that include year effects and initial conditions, and that use the more non-parametric set of instruments (I HI1 , the share of Brazil in exports in 1998 times year effects, and I EXP , the weighted average exchange rate, de�ned in equations (10) and (12), respectively). In Table 8, we con�rm our conclusion that exporting per se does not really matter but that exporting to high-income countries does. In fact, we �nd that under the more stringent de�nition of high-income 24 destinations the impacts are close to three times larger than under De�nition I. For instance, in the model of wages controlling for log sales, the coefficient on HI is 0.836 instead of 0.313. This result is consistent with our hypothesis since we expect larger differences in α and τ , and thus in skill utilization, the higher the income of the high-income destination group. Our conclusions are robust to results (not shown) using the alternative set of instruments given by (11), the share of Brazil in export in 1998 times the Brazilian exchange rate, and (12). 4 Channels In this section, we set out to uncover some of the channels behind our “export destinations matter� result. We want to shed some light on the mechanisms by which Argentine �rms that became more oriented towards high-income destinations utilized more skills. 4.1 Skill Upgrading Within Labor Categories We begin with a simple extension of our baseline regression model. We re-estimate the wage speci�cation with the addition of the share of non-production workers as a regressor. Our objective is to test whether HI remains statistically signi�cant after controlling for the share of non-production workers. We include, as before, both HI and EXP as explanatory variables, and we instrument them with I HI1 and I EXP . Results are reported in Table 9 (similar results are obtained when we use I HI2 , in place of I HI1 , and I EXP as instruments). We �nd that export intensity, EXP , is never signi�cant, while HI, exports to high-income countries, always is. As expected, the share of non-production workers is positively associated with the average �rm wage since on average non-production workers are more skilled than production workers. In addition, the coefficient of HI is smaller than in all previous regressions. Taken together, these results indicate that part of the impact of HI on average wages effectively work through increases in the share of non-production workers but that there are also other channels playing a role. One plausible mechanism is skill upgrading within labor categories. With two labor categories, the average wage will (w) can be written as (13) w = wnp lnp + wp lp , where wnp and wp are the average wages paid to non-production and production workers, and 25 lnp and lp are their shares in total �rm employment. In this formulation, our �ndings so far revealed that �rms exporting to high-income countries pay higher wages w, hire a higher share lnp of non-production workers, and pay higher wages w conditional on the shares lnp . This last result is consistent with a scenario where �rms engage in skill upgrading within skill categories and possibly utilize better (that is, more skilled) non-production and production workers (in at least one of the two categories). To further clarify this idea, assume that there are two categories of workers, say semiskilled and skilled, within the non-production worker category, with the wages of the skilled workers being higher than the wages of the semiskilled workers. A higher lnp associated with high-income exports can take place because of higher levels of skilled workers or of higher levels of semiskilled workers. It follows that, if the mechanism operates via more skilled workers, then average wages will be even higher. This channel is consistent with the within-category skill upgrading uncovered by Verhoogen (2008).24 4.2 Quality In our theory, the quality upgrading and the required services mechanisms played major roles. If these channels are valid, we argue that they should be much stronger in sectors with higher scope for quality/services upgrades. To test this, we estimate our IV regressions after splitting the sample according to the variance of the unit values in exports at the industry level (a measure of the degree of product differentiation in the sector). We calculated unit values using industry-level bilateral trade data from COMTRADE. First, we matched the COMTRADE trade data to the industrial classi�cation system from the �rm survey, the ISIC Revision 2 at the 3 digit level. Second, for each 3-digit industry, we computed the variance in export unit values from all pairwise combinations of countries of origin and destination that report to COMTRADE, after trimming outliers. Finally, we classi�ed industries as “High Variance,� if the variance of their unit values is above the 75th percentile, or as “Low Variance� otherwise. Results are displayed in Table 10. We report the estimates of the IV models that use the more non-parametric instruments (year dummies times initial shares (equation 10) and the average exchange rate (equation 12)) and that control for year effects and �rm initial conditions. The results are however robust to all other speci�cations presented above. In Panel A, we use log average wages as measures of skill utilization. We �nd that exporters to high-income countries pay higher wages 24 Note that, as we discussed in the theoretical model, we cannot rule out other possible mechanisms like pro�t-sharing (fair wages), labor turnover, or scale economies. 26 in industries in which there is more vertical differentiation but not in industries in which products are vertically more homogeneous.25 In Panel B, we use instead the share of non-production workers (number of employees). In this case, we �nd that HI impacts the share of non-production workers in both High-variance and Low-variance sectors.26 These �ndings provide support to both Matsuyama (2007) and Verhoogen (2008). On the one hand, the fact that high-income exporters pay higher wages only in sectors with a higher scope for quality differentiation is clearly consistent with Verhoogen’s mechanism. On the other hand, the fact that high-income exporters hire a higher share of non-production workers is consistent with “required services� that are independent of the degree of vertical differentiation, as in Matsuyama (2007). Also, taken together, both results suggest that skill upgrading due to vertical differentiation is materialized via skill upgrading within categories (as found by Verhoogen (2008)), while Matsuyama’s technical services require a higher share of non-production workers. In terms of equation 13, our �ndings can be interpreted as follows. Matsuyama implies higher lnp associated with high-income exports in all sectors. In turn, we only observe higher w in high-quality sectors. This means that the quality mechanism in high-quality sectors operates via skilled upgrading, that is via more skilled workers vis-‘a-vis semiskilled workers among the non-production workers, lnp . The “required services� of Matsuyama, in contrast, are met with semiskilled workers, rather than skilled workers, within non-production workers. 5 Conclusions In this paper, we elaborate upon a theory linking export destinations and skill utilization in developing countries. We provide a uni�ed theoretical framework to study the behavior of �rms that export to high-income countries in terms of the utilization of skilled labor. In our framework, exporters to high-income destinations hire a higher level of non-production workers for two reasons. 25 We have experimented with several cutoffs for high and low variance industries. We do not �nd different impacts between high and low variance industries if we de�ne high variance industries as those above the mean or the median. 26 For robustness, we also split the sample using the length of the quality ladder introduced by Khandelwal (2009). He calculates an index based on the estimation of demand equations that incorporate a valuation for quality. We de�ned sectors with “long� quality ladders if his index is above the mean, and with “short� ladders in the opposite case. Overall, our results are similar to those in Table 10. When the dependent variable is log average wages, we �nd a strong, positive, and statistically signi�cant impact of exporting to high-income export destinations (HI) in industries with long ladders. In industries with short ladders, the estimates of δ2 are smaller and they are not statistically signi�cant. When the dependent variable is the share of non-production workers, we �nd positive impacts of HI on the share of non-production workers in both long- and short-ladder industries. The point estimates are similar in both cases (and also comparable to those in Table 10), but they are only marginally signi�cant. These results can be found in our online Appendix at http://faculty.udesa.edu.ar/ibrambilla/papers/exportdestinations_appendix.pdf. 27 First, exporting to high-income countries requires quality upgrades, which are skill-intensive. Second, there are required services associated with exporting to high-income countries, and these activities are also intensive in skills. Our model introduces multi-dimensional �rm heterogeneity in order to explain both the decision to export as well as the decision to export to high-income countries. This heterogeneity is due to differences in the efficiency in the use of skilled and unskilled labor across �rms. We provide empirical evidence using a panel of Argentine manufacturing �rms that we matched with customs information on exports and export destinations at the �rm level. The available data cover the 1998-2000 period and thus span the Brazilian devaluation of 1999. This provides a useful source of identi�cation because we can use the exogenous changes in exports and in export destinations brought about by the Brazilian devaluation to explore whether �rms choose the skill composition of their workforce based on the destination of their exports. The empirical models consistently �nd that exporting to high-income countries induces �rms to hire more skilled workers, but exporting per se does not. The reason is that the domestic markets in Argentina are similar to export markets in middle-income countries and thus it is only possible to observe differences in �rm’s outcomes for �rms specializing in exporting to high-income countries. We also provide evidence in support of the two mechanisms outlined in our model, namely quality upgrades and required services. Our contribution lies in identifying, empirically and theoretically, mechanisms that explain how the “act of exporting� to different destinations affects the behavior of �rms. Our results can shed light on the nature of this behavior and, in turn, this may prove useful in current research efforts to understand factors driving �rm choices of exporting and of exporting to different markets. Appendices Appendix 1: Distribution of Firms by Industry Table A1 reports the number of �rms by 2-digit industry in 1998. This information gives a sense of the type of industries involved in our analysis. Among a total of 901 manufacturing �rms, the largest number of units are in Food & Beverages (139), followed by Chemical Products (83), Textiles (68), Rubber & Plastic (67) and Metal Products (66). In contrast, there are very few �rms in Coke & Re�ned Petroleum Products, Office, Accounting & Computing Machinery, or Radio, TV & Communication Equipment. The survey does not cover �rms that produce primary products like agricultural commodities. 28 Appendix 2: Export Destinations De�nitions Table A2 lists the World Bank country classi�cation. Appendix 3: One Plant Firms As we mentioned during the description of the data, the ENI �rm survey collects data at the plant level while the Customs data collect export information at the �rm level. We deal with this issue by aggregating the plant-level data. To see if this has any impact on our estimates, we re-did all the regressions reported in the text for the sample of one-plant �rms. Since 86 percent of �rms in the ENI survey are one-plant �rms, we do not expect sizeable differences in our estimates. In Table A3, we report results for log average shares and for the share of non-production workers. None of our conclusions are affected. Appendix 4: Alternative Set of Instruments In this appendix, we test the robustness of our results by using an alternative set of instruments for export density, EXP , the share of exports on total sales. In Section 3, we built a measure of the average exchange rate faced by a �rm by weighing the exchange rate of each country partner with the initial export share in sales in 1998. In what follows, we replace this instrument with a similar one that only exploits the exogenous variation caused by the Brazilian devaluation. Concretely, we de�ne EXP BRA (14) Iit 1 = φt ∗ ψi98 , and EXP BRA (15) Iit 2 = erateit ∗ ψi98 . EXP The instrument Iit 1 is analogous to I HI1 ; we interact the year effects φt with the initial share BRA of exports to Brazil on total sales, ψ98 . To construct I EXP2 , we interact those shares with the Brazilian exchange rate so that this instrument works in the same way as I EXP does (but only taking into account the Brazilian devaluation).27 Note that I EXP1 and I EXP2 are a measure of the “scope for retrenchment into local markets.� Firms with a larger pre-shock share of exports to Brazil on total sales had more possibilities to divert sales into the domestic markets (compared to other �rms oriented towards non-Brazil markets). We expect a negative association between EXP and I EXP1 or I EXP2 in the �rst stage. In addition, we expect these instruments to be negatively correlated with HI because a higher scope for retrenchment into local markets limit the scope to divert exports to high income. In contrast, we expect I HI1 and I HI2 to be positively correlated with export intensity EXP because a higher scope for export switching to high-income countries allow �rm exports to remain high relative to sales (conditional of the other controls and instruments). In principle, we have four possible combinations of instruments. To simplify the exposition, we only consider two combinations, one where we use time dummies to build the instruments for both HI and EXP , and another where we use the Brazilian exchange rates. The results are reported in Tables A4 (for the case of instruments based on year dummy variables, equation (14)) and Table A5 (for the case of instruments based on exchange rates, equation (15)). In Panel A, we report IV 27 Note the difference with λBRA in (10) or (11), which is the share of exports to Brazil on total exports. That is, i98 ai98 =exports to Brazil/total exports but bi98 =exports to Brazil/sales. 29 estimates using both log average wages and the share of non-production workers (in numbers) as measures of skill utilization. In all these regressions, our conclusions are unchanged: only exporting to high-income destinations matters. The point values of the estimates are very similar to those reported in the main text. For instance, in column 3 of Table A4, the IV coefficient of HI is 0.311 a (vis-`-vis 0.313 in Table 4). In column 6 of Table A4, the IV coefficient of HI is the non-production workers share equation is 0.124 (vis-a-vis 0.121). In Panels B1 and B2 of Tables A4 and A5, we also report the �rst-stage results. Overall, the instruments work very well, have good explanatory power and are highly statistically signi�cant. There is no apparent risk of weak instrumentation. Appendix 5: Additional Results In the paper, we only show a set of the most important results to illustrate the �ndings. We have prepared a note with a full set of results, which can be found at http://faculty.udesa.edu.ar/ ibrambilla/papers/exportdestinations_appendix.pdf. This Note includes tables with results using instruments I HI2 and I EXP for the following models: i) wage regressions controlling for share of non-production workers; ii) High-variance / Low-Variance regressions; iii) Quality ladder regressions. We also include tables with results using the alternative instruments described in Appendix 4 for the following models: i) wage regressions conditional on share of non-production workers; ii) High-variance/Low-variance regressions; iii) Quality Ladders. Finally, we include a full set of results with various instruments using De�nition II of high-income destinations. References Amity, M. and D. Davis (2008). “Trade, Firms, and Wages: Theory and Evidence,� NBER Working Paper No 14106. Arkolakis, K. (2009). “Market Penetration Costs and the New Consumers Margin in International Trade,� mimeo Yale University. Arkolakis, K., S. Demidova, P. Klenow, and A. Rodriguez-Clare (2008). “The Gains from Trade with Endogenous Variety,� American Economic Review Papers and Proceedings, 98 (4), pp. 444-450. Arellano, M. (2003). Panel Data Econometrics, Advanced Texts in Econometrics, Oxford University Press, New York. Bernard, A. and J. Jensen (1995). “Exporters, Jobs, and Wages in U.S. Manufacturing: 1976-1987,� Brooking Papers on Economic Activity: Microeconomics, pp. 67-119. Bernard, A. and J. Jensen (1999). “Exceptional Exporter Performance: Cause, Effect, or Both?,� Journal of International Economics, 47, pp. 1-25. 30 Bernard, A.B. and Jensen, J.B. (2004). “Why Some Firms Export,� Review of Economics and Statistics 86, pp. 561-569. Bernard, A., S. Redding, and P. Schott (2009). “Multi-Product Firms and Trade Liberalization,� mimeo Yale School of Management (revised version of NBER Working Paper 12782). Bernard, A., B. Jensen, S. Redding, and P. Schott (2007). “Firms in International Trade,� Journal of Economic Perspectives,� Volume 21, Number 3, Summer 2007, pp. 105–130. Bustos, P. (2009). “Trade Liberalization, Exports and Technology Upgrading: Evidence on the impact of MERCOSUR on Argentinean Firms,� forthcoming, American Economic Review. Chaney, T. (2008). “Distorted Gravity: The Intensive and Extensive Margins of International Trade,� American Economic Review 98:4, pp. 1707–1721. Clerides, S., S. Lach, and J. Tybout (1998). “Is Learning by Exporting Important? Micro-dynamic Evidence from Colombia, Mexico, and Morocco,� Quarterly Journal of Economics, 108(3), pp. 903-947. De Loecker, J. (2007). “Do Exports Generate Higher Productivity? Evidence from Slovenia,� Journal of International Economics 73, pp. 69–98. Egger H. and U. Kreickemeier (2009). “Firm Heterogeneity And The Labor Market Effects of Trade Liberalization,� International Economic Review, vol. 50(1), pp. 187-216. Fajgelbaum, P., G. Grossman and E. Helpman (2009). “Income Distribution, Product Quality and International Trade,� Princeton University. Feenstra, R. and G. Hanson (1996). “Globalization, Outsourcing, and Wage Inequality, � American Economic Review, vol. 86(2), pp. 240–45 Galiani, S. and G. Porto (2009). “Trends in Tariff Reforms and Trends in the Structure of Wages,� The Review of Economics and Statistics, forthcoming. Grossman, G. and E. Rossi-Hansberg (2008). “Trading Tasks: A Simple Theory of Offshoring,� American Economic Review, vol. 98(5), pp.1978–1997 Grossman, G. and E. Rossi-Hansberg (2009). “Task Trade between Similar Countries,� mimeo Princeton University 31 Hallak, J. C. (2006). “Product Quality and the Direction of Trade,� Journal of International Economics, 62, pp. 238-265. Hallak, J. C. (2008). “A Product-Quality View of the Linder Hypothesis,� forthcoming in The Review of Economics and Statistics. Hallak, J.C. and P. Schott (2008). “Estimating Cross-Country Differences in Product Quality,� Yale SOM. Hallak, J.C. and J. Sivadasan (2009). “Firms’ Exporting Behavior under Quality Constraints,� International Policy Center Working Paper No. 88, University of Michigan. Hausmann, R., J. Hwang, and D. Rodrik (2007). “It Is Not How Much But What You Export That Matters,� Journal of Economic Growth, 12, pp. 1-25. Hummels, D. and A. Skiba (2004). “Shipping the Good Apples Out? An Empirical Con�rmation of the Alchian-Allen Conjecture,� Journal of Political Economy, 102, pp. 1384–1402. Hummels, D. and P. Klenow (2005). “The Variety and Quality of a Nation’s Exports,� American Economic Review, 95, pp. 704–723. Khandelwal, A. (2009). “The Long and Short (of) Quality Ladders,� forthcoming, Review of Economic Studies. Kugler, M. and E. Verhoogen (2008). “The Quality-Complementarity Hypothesis: Theory and New Evidence from Colombia,� mimeo Columbia. Manasse, P. and A. Turrini (2001). “Trade, Wages, and ‘Superstars’,� Journal of International Economics 54, pp. 97–117. Matsuyama, K. (2007). “Beyond Icebergs: Towards A Theory of Biased Globalization,� The Review of Economic Studies, 74 , pp. 237-253. Melitz, M. (2003). “The Impact of Trade on Intra-industry Reallocations and Aggregate Industry Productivity,� Econometrica, 71, pp. 1695-1725. Milner, C. and Tandrayen, V. (2007). “The Impact of Exporting and Export Destination on Manufacturing Wages: Evidence for Sub-Saharan Africa,� Review of Development Economics, 11(1), pp. 13–30. 32 Park, A., D. Yang, X. Shi, and Y. Jiang (2008). “Exporting and Firm Performance: Chinese Exporters and the Asian Financial Crisis,� Review of Economics and Statistics, forthcoming. Pavcnik, N. (2002). “Trade Liberalization, Exit, and Productivity Improvements: Evidence from Chilean Plants,� Review of Economic Studies, Vol. 69, No. 1, 245-276. Revenga, A. (1992). “Exporting Jobs? The Impact of Import Competition on Employment and Wages in U.S. Manufacturing,� Quarterly Journal of Economics, 107, pp. 255–284. Schank, T., C. Schnabel, and J. Wagner (2007). “Do exporters really pay higher wages? First evidence from German linked employer-employee data,� Journal of International Economics, vol. 72(1), pp. 52-74. Stock, J. and D. Staiger (1997). “Instrumental Variables Regression with Weak Instruments,� Econometrica, 65, No 3, pp. 557-586. Verhoogen, E. (2008). “Trade, Quality Upgrading, and Wage Inequality in the Mexican Manufacturing Sector,� Quarterly Journal of Economics, Vol 123, No. 2, pp. 489-530. Yeaple (2005). “A Simple Model of Firm Heterogeneity, International Trade, and Wages,� Journal of International Economics 65, pp. 1– 20. 33 Table 1 Descriptive Statistics from Firm Survey (ENI) and Customs Records Argentina 1998-2000 All Firms Exporters High-income exporters I II (1) (2) (3) (4) Panel A Exported in a given year 0.59 Exported during sample period 0.68 Exports/Sales 0.08 0.13 0.15 0.22 Number of destinations 3.3 4.9 5.4 7.8 Observations 2544 1499 1307 680 Panel B Number of workers 89.7 1.22∗∗∗ 0.39∗∗∗ 0.43∗∗∗ ( 0.04 ) ( 0.07 ) (0.05) Annual sales in 100,000 USD 8.04 1.73∗∗∗ 0.54∗∗∗ 0.54∗∗∗ ( 0.05 ) ( 0.09 ) (0.06) Average annual wage in USD 12,154 0.48∗∗∗ 0.12∗∗∗ 0.15∗∗∗ ( 0.02 ) ( 0.03 ) (0.03) Share of non-production workers 0.26 0.05∗∗∗ 0.03∗∗∗ 0.014 (Number of workers) ( 0.01 ) ( 0.01 ) (0.01) Share of non-production workers 0.26 0.05∗∗∗ 0.03∗∗∗ 0.02∗∗ (Hours worked) ( 0.01 ) ( 0.01 ) (0.01) Source: Own calculations based on �rm data from the National Industrial Survey (ENI) and customs records. (1): Average number of workers, average annual sales, average annual wage, average share of non-production workers (number of workers and hours worked). (2): Difference in means in log workers, log sales, log wage and share of non-production workers between exporters and non-exporters, controlling for industry and year. (3): Difference in means between �rms that export to at least one high-income destination and other exporters (conditional on exporting), controlling for industry and year. (4): Difference in means between �rms that export to at least one high-income destination (de�nition II) and other exporters (conditional on exporting to high-income countries, de�nition I), controlling for industry and year. De�nition I includes high-income and upper-middle-income countries. De�nition II includes only high-income countries. 34 Table 2 Exports, Export Destinations and Skill Utilization in the Cross Section OLS Estimates Log Average Wage Share of non-production Workers (1) (2) (3) (4) (5) (6) ∗∗∗ ∗∗∗ Exports/Sales (EXP ) 0.692 − 0.513 0.0410 − 0.010 (0.110) (0.113) (0.031) (0.032) High Income Exports (HI) − 0.340∗∗∗ 0.291∗∗∗ − 0.052∗∗∗ 0.051∗∗∗ (0.039) (0.291) (0.012) (0.013) Observations 2544 2544 2544 2544 2544 2544 R-squared 0.219 0.238 0.256 0.257 0.266 0.266 All regressions include year and industry effects. Standard errors are clustered at the �rm level. Signi�cance at the 1 percent, 5 percent and 10 percent level is denoted by ***, ** and *. The share of non-production workers is measured as the number of non-production workers relative to the total number of employees. Table 3 Main Countries of Destination of Argentine Manufacturing Exports 1998 1999 2000 Value Share Value Share Value Share Brazil 5568.5 0.36 3858.3 0.28 4363.6 0.31 United States 1550.9 0.10 1822.7 0.13 2187.1 0.15 Chile 959.8 0.06 950.4 0.07 1190.0 0.08 Uruguay 654.0 0.04 638.8 0.05 608.6 0.04 Paraguay 491.5 0.03 441.3 0.03 460.1 0.03 Europe 2025.3 0.13 2037.7 0.15 2014.8 0.14 TOTAL 15259.1 1 13716.0 1 14155.9 1 High Income I 6512.3 0.43 6840.5 0.50 7265.8 0.51 High Income II 4237.3 0.28 4624.7 0.34 4872.3 0.34 Source: UN COMTRADE. Manufacturing sector only. Values in constant 1998 millions of dollars. High Income I: countries classi�ed by the World Bank as high income (OECD and non-OECD) and upper middle income. High Income II: countries classi�ed by the World Bank as high income (OECD and non-OECD). 35 Table 4 Exports, Export Destinations, and Skills. Wage Regressions Dummy Instruments (1) (2) (3) (4) (5) Panel A: Second Stage Exports/Sales (EXP ) –0.186 –0.191 –0.324 –0.273 –0.235 (0.475) (0.518) (0.556) (0.527) (0.544) High Income Exports (HI) 0.277∗∗∗ 0.306∗∗∗ 0.321∗∗∗ 0.338∗∗∗ 0.313∗∗∗ (0.098) (0.098) (0.106) (0.108) (0.104) Log Sales − − − − 0.058∗∗∗ (0.020) Panel B1: First Stage (HI) Share BRA in exports * 1999 (I HI1 ) 0.213∗∗∗ 0.230∗∗∗ 0.231∗∗∗ 0.229∗∗∗ 0.231∗∗∗ (0.035) (0.039) (0.041) (0.040) (0.041) Share BRA in exports * 2000 (I HI1 ) 0.241∗∗∗ 0.250∗∗∗ 0.245∗∗∗ 0.249∗∗∗ 0.245∗∗∗ (0.038) (0.043) (0.045) (0.043) (0.045) Average Exchange Rate (I EXP ) 0.986∗∗∗ 0.970∗∗∗ 0.980∗∗∗ 0.973∗∗∗ 0.992∗∗∗ (0.302) (0.304) (0.308) (0.306) (0.309) Log Sales − − − − 0.010 (0.017) R-squared 0.034 0.035 0.035 0.035 0.035 p-value 0.000 0.000 0.000 0.000 0.000 Panel B2: First Stage (EXP ) Share BRA in exports * 1999 (I HI1 ) 0.012 0.013 0.011 0.010 0.011 (0.011) (0.011) (0.012) (0.012) (0.012) Share BRA in exports * 2000 (I HI1 ) 0.037∗∗∗ 0.031∗∗ 0.029∗∗ 0.029∗∗ 0.029∗ (0.013) (0.014) (0.014) (0.014) (0.014) Average Exchange Rate (I EXP ) 0.600∗∗∗ 0.612∗∗∗ 0.619∗∗∗ 0.618∗∗∗ 0.626∗∗∗ (0.229) (0.231) (0.232) (0.232) (0.232) Log Sales − − − − 0.006 (0.009) R-squared 0.029 0.031 0.031 0.031 0.033 p-value 0.001 0.007 0.011 0.006 0.011 Year Effects − Yes Yes Yes Yes Initial Conditions * Year Effects − − Yes − Yes Initial Conditions * Exchange Rate − − − Yes − Number of �rms 901 901 901 901 901 Observations 2544 2544 2544 2544 2544 Dependent variable in second stage: Log Average Wage (Panel A). Dependent variables in �rst stage: Exports to high income destinations over Total value of exports (HI), in Panel B1; and Total value of exports over Total value of sales (EXP ) in Panel B2. All regressions include �rm �xed effects. Dummy instruments are used in all regressions. Bootstrapped SE in parenthesis. Signi�cance at the 1 percent, 5 percent and 10 percent level is denoted by ***, ** and *. 36 Table 5 Exports, Export Destinations, and Skills. Wage Regressions Exchange Rate Instruments (1) (2) (3) (4) (5) Panel A: Second Stage Exports/Sales (EXP ) –0.258 –0.370 –0.356 –0.442 –0.232 (0.645) (0.620) (0.613) (0.627) (0.580) High Income Exports (HI) 0.270∗∗∗ 0.282∗∗∗ 0.316∗∗∗ 0.311∗∗∗ 0.314∗∗∗ (0.109) (0.105) (0.109) (0.111) (0.106) Log Sales − − − − 0.058∗∗∗ (0.020) Panel B1: First Stage (HI) Share BRA in exports * erate (I HI2 ) 0.364∗∗∗ 0.393∗∗∗ 0.395∗∗∗ 0.392∗∗∗ 0.395∗∗∗ (0.057) (0.063) (0.065) (0.065) (0.065) Average Exchange Rate (I EXP ) 0.814∗∗∗ 0.871∗∗∗ 0.905∗∗∗ 0.873∗∗∗ 0.919∗∗∗ (0.287) (0.297) (0.305) (0.300) (0.306) Log Sales − − − − 0.012 (0.017) R-squared 0.310 0.032 0.033 0.032 0.33 p-value 0.000 0.000 0.000 0.000 0.000 Panel B2: First Stage (EXP ) Share BRA in exports * erate (I HI2 ) 0.028 0.029 0.025 0.025 0.026 (0.019) (0.020) (0.021) (0.020) (0.020) Average Exchange Rate (I EXP ) 0.534∗∗ 0.580∗∗ 0.594∗∗ 0.587∗∗ 0.602∗∗∗ (0.225) (0.230) (0.232) (0.230) (0.231) Log Sales − − − − 0.006 (0.009) R-squared 0.020 0.027 0.028 0.027 0.029 p-value 0.059 0.041 0.037 0.039 0.034 Year Effects − Yes Yes Yes Yes Initial Conditions * Year Effects − − Yes − Yes Initial Conditions * Exchange Rate − − − Yes − Number of �rms 901 901 901 901 901 Observations 2544 2544 2544 2544 2544 Dependent variable in second stage: Log Average Wage (Panel A). Dependent variables in �rst stage: Exports to high income destinations over Total value of exports (HI), in Panel B1; and Total value of exports over Total value of sales (EXP ) in Panel B2. All regressions include �rm �xed effects. Exchange rate instruments are used in all regressions. Bootstrapped SE in parenthesis. Signi�cance at the 1 percent, 5 percent and 10 percent level is denoted by ***, ** and *. 37 Table 6 Exports, Export Destinations, and Skills. Share of Non-Production Workers Dummy Instruments (1) (2) (3) (4) (5) Panel A: Workers Exports/Sales (EXP ) –0.109 –0.015 0.011 0.011 0.007 (0.150) (0.165) (0.164) (0.166) (0.162) High Income Exports (HI) 0.174∗∗∗ 0.130∗∗∗ 0.120∗∗∗ 0.120∗∗∗ 0.121∗∗∗ (0.042) (0.041) (0.041) (0.041) (0.042) Log Sales − − − − -0.003 0.007 Panel B: Hours Exports/Sales (EXP ) –0.126 –0.001 0.026 0.030 0.022 (0.162) (0.169) (0.175) (0.175) (0.171) High Income Exports (HI) 0.171∗∗∗ 0.105∗∗∗ 0.092∗∗ 0.093∗∗ 0.093∗∗ (0.043) (0.041) (0.042) (0.040) (0.042) Log Sales − − − − –0.004 (0.007) Year Effects − Yes Yes Yes Yes Initial Conditions * Year Effects − − Yes − Yes Initial Conditions * Exchange Rate − − − Yes − Observations 2544 2544 2544 2544 2544 Number of Firms 901 901 901 901 901 Dependent variables: number of non-production workers over total number of workers (Panel A) and hours worked by non-production workers over total number of hours worked by all employees (Panel B). All regressions include �rm �xed effects and year effects. Dummy instruments are used in all regressions. First stage regressions are the same as in Table 4. Bootstrapped SE in parenthesis. Signi�cance at the 1 percent, 5 percent and 10 percent level is denoted by ***, ** and *. 38 Table 7 Exports, Export Destinations, and Skills. Share of Non-Production Workers Exchange Rate Instruments (1) (2) (3) (4) (5) Panel A: Workers Exports/Sales (EXP ) –0.078 0.018 0.042 0.041 0.035 (0.190) (0.180) (0.179) (0.179) (0.177) High Income Exports (HI) 0.177∗∗∗ 0.134∗∗∗ 0.125∗∗∗ 0.125∗∗∗ 0.125∗∗∗ (0.044) (0.041) (0.043) (0.043) (0.043) Log Sales –0.003 (0.007) Panel B: Hours Exports/Sales (EXP ) –0.086 0.071 0.108 0.099 0.098 (0.199) (0.196) (0.197) (0.198) (0.194) High Income Exports (HI) 0.175∗∗∗ 0.115∗∗∗ 0.104∗∗ 0.104∗∗ 0.104∗∗ (0.044) (0.042) (0.043) (0.043) (0.043) Log Sales –0.005 (0.007) Year Effects − Yes Yes Yes Yes Initial Conditions * Year Effects − − Yes − Yes Initial Conditions * Exchange Rate − − − Yes − Observations 2544 2544 2544 2544 2544 Number of Firms 901 901 901 901 901 Dependent variables: number of non-production workers over total number of workers (Panel A) and hours worked by non-production workers over total number of hours worked by all employees (Panel B). All regressions include �rm �xed effects and year effects. Exchange rate instruments are used in all regressions. First stage regressions are the same as in Table 4. Bootstrapped SE in parenthesis. Signi�cance at the 1 percent, 5 percent and 10 percent level is denoted by ***, ** and *. 39 Table 8 Alternative De�nition of High Income Exports Log Average Wage Share of non-prod Workers (1) (2) (3) (4) (5) (6) Exports/Sales (EXP ) −0.339 −0.264 −0.239 0.010 0.013 0.008 (0.553) (0.532) (0.542) (0.169) (0.163) (0.165) High Income Exports II (HI) 0.894∗∗∗ 0.968∗∗∗ 0.861∗∗∗ 0.324∗∗ 0.346∗∗∗ 0.325∗∗ (0.322) (0.332) (0.314) (0.130) (0.132) (0.130) Log Sales − − 0.061∗∗∗ − − -0.002 (0.022) (0.008) Initial Conditions * Year Effects Yes − Yes Yes − Yes Initial Conditions * Exchange Rate − Yes − − Yes − Observations 2544 2544 2544 2544 2544 2544 Firms 901 901 901 901 901 901 High Income Exports II are de�ned as the share of total value of exports that is shipped to countries classi�ed by the World Bank as High Income OECD and non-OECD (it excludes upper-middle income countries). Dependent variables: Log average wage in columns (1), (2), (3); Number of non-production workers over total number of workers in columns (4), (5), (6). All regressions include �rm �xed effects and year effects. Dummy instruments are used in all regressions. Bootstrapped SE in parenthesis. Signi�cance at the 1 percent, 5 percent and 10 percent level is denoted by ***, ** and *. Table 9 Exports, Export Destination and Skills Wage Regressions Controlling for Share of Non-Production Workers. (1) (2) (3) Exports/Sales (EXP ) –0.330 –0.278 –0.239 (0.517) (0.501) (0.522) High Income Exports (HI) 0.265∗∗∗ 0.281∗∗∗ 0.256∗∗∗ (0.101) (0.103) (0.098) Share of non-prod workers 0.469∗∗∗ 0.471∗∗∗ 0.471∗∗∗ (0.088) (0.089) (0.088) Log Sales − − 0.059∗∗∗ (0.019) Initial Conditions * Year Effects Yes − Yes Initial Conditions * Exchange Rate − Yes − Observations 2544 2544 2544 Firms 901 901 901 Dependent variable: Log average wage. Share of non-production workers is measured as the number of non-production workers over the total number of workers. All regressions include �rm �xed effects and year effects. Dummy instruments are used in all regressions. Bootstrapped SE in parenthesis. Signi�cance at the 1 percent, 5 percent and 10 percent level is denoted by ***, ** and *. 40 Table 10 Scope for Vertical Differentiation. Variance in Unit Values High Variance Industries Low Variance Industries (1) (2) (3) (4) (5) (6) Panel A: Average Wages Exports/Sales (EXP ) –0.301 –0.301 –0.268 –0.296 –0.196 –0.138 (0.908) (0.921) (0.887) (0.459) (0.451) (0.439) High Income Exports (HI) 0.461∗∗ 0.476∗∗ 0.435∗∗ 0.211 0.230 0.214 (0.187) (0.188) (0.194) (0.140) (0.141) (0.137) Log Sales − − 0.050 − − 0.062∗∗ (0.034) (0.029) Observations 973 973 973 1506 1506 1506 Firms 344 344 344 536 536 536 Panel B: Non-Production Workers Exports/Sales (EXP ) 0.077 0.080 0.073 –0.026 –0.041 –0.028 (0.234) (0.247) (0.237) (0.216) (0.211) (0.203) High Income Exports (HI) 0.121∗ 0.122∗ 0.124∗ 0.110∗∗ 0.107∗∗ 0.109∗∗ (0.072) (0.072) (0.072) (0.054) (0.055) (0.055) Log Sales − − -0.005 − − –0.002 (0.015) (0.009) Observations 973 973 973 1506 1506 1506 Firms 344 344 344 536 536 536 Initial Conditions * Year Effects Yes − Yes Yes − Yes Initial Conditions * Exchange Rate − Yes − − Yes − Dependent variable: Log average wage (Panel A) and Share of non-production workers (Panel B). High (Low) variance: 3-digit ISIC industries with variance in export unit values above (below) the 75th percentile. All regressions include �rm �xed effects and year effects. Dummy instruments are used in all regressions. Bootstrapped SE in parenthesis. Signi�cance at the 1 percent, 5 percent and 10 percent level is denoted by ***, ** and *. 41 Table A1 Distribution of Firms by Industry All Exporters Non- �rms Exporters Food and beverages 139 72 67 Textiles 68 38 30 Apparel 17 7 10 Leather and leather products 22 10 12 Wood, cork and straw products 20 5 15 Paper and paper products 31 20 11 Publishing, printing, media 27 10 17 Coke and re�ned petroleum products 4 4 Chemicals and chemical products 83 73 10 Rubber and plastics products 67 42 25 Other non-metallic mineral products 61 29 32 Basic metals 34 20 14 Metal products 66 35 31 Machinery and equipment n.e.c. 82 62 20 Office, accounting and computing machinery 1 1 Electrical machinery 56 37 19 Radio, TV and communication equipment 5 3 2 Medical, precision and optical instruments 13 11 2 Motor vehicles 44 30 14 Other transport equipment 14 7 7 Furniture; Other 47 27 20 Total 901 543 358 Source: Own calculations based on �rm data from the Encuesta Nacional Industrial (ENI) or National Industrial Survey. Industries are classi�ed according to ISIC Revision 2. In this table we aggregate industries at the 2-digit level. 42 Table A2 World Bank Country Classi�cation High income OECD: Australia, Austria, Belgium, Canada, Denmark, Finland, France, Germany, Greece, Iceland, Italy, Japan, Korea, Luxembourg, Netherlands, Norway, New Zealand, Portugal, Spain, Sweden, Switzerland, United Kingdom, United States High income non-OECD: Bahrain, Bahamas, Barbados, Cyprus, Hong Kong, Israel, Kuwait, Malta, Puerto Rico, Qatar, Singapore, United Arab Emirates Upper-middle income: Chile, Costa Rica, Croatia, Czech Republic, Dominica, Gabon, Grenada, Hungary, Lebanon, Lithuania, Mexico, Mauritius, Malaysia, Panama, Poland, Saudi Arabia, Seychelles, St. Lucia, Trinidad and Tobago, Uruguay, Venezuela Low-middle income: Algeria, Bolivia, Brazil, Bulgaria, China, Colombia, Cuba, Dominican Republic, Ecuador, Egypt, El Salvador, Guatemala, Guyana, Honduras, Iran, Jamaica, Jordan, Morocco, Paraguay, Peru, Phillipines, Russia, Saint Vincent and the Grenadines, South Africa, Sri Lanka, Suriname, Syria, Thailand, Turkmenistan, Tunisia, Turkey, Ukraine Low income: Angola, Benin, Bangladesh, Cote d’Ivorie, Comoros, Democratic People’s Republic of Korea, Democratic Republic of Congo, Ethiopia, Haiti, India, Indonesia, Kenya, Myanmar, Mozambique, Nigeria, Nicaragua, Pakistan, Papua New Guinea, Tanzania, Uganda, Vietnam, Yemen, Zimbabwe List includes countries with positive exports in the �rm-level data. 43 Table A3 One Plant Firms Only Log Average Wage Share non-prod Workers (1) (2) (3) (4) (5) (6) Exports/Sales (EXP ) –0.391 –0.302 –0.339 –0.142 –0.146 –0.15 (0.55) (0.524) (0.526) (0.155) (0.157) (0.150) High Income Exports (HI) 0.324∗∗ 0.333∗∗ 0.324∗∗ 0.097∗∗ 0.098∗∗ 0.097∗∗ (0.127) (0.134) (0.127) (0.047) (0.047) (0.047) Log Sales − − 0.045∗∗ − − –0.005 (0.021) (0.009) Initial Conditions * Year Effects Yes − Yes Yes − Yes Initial Conditions * Exchange Rate − Yes − − Yes − Observations 2117 2117 2117 2117 2117 2117 Number of Firms 750 750 750 750 750 750 Firms that report owning more than one plant are dropped from the sample. Dependent variables: Log average wage in columns (1)-(3); number of non-production workers over total number of workers in columns (4)-(6). All regressions include �rm �xed effects and year effects. Dummy instruments are used in all regressions. Bootstrapped SE in parenthesis. Signi�cance at the 1 percent, 5 percent and 10 percent level is denoted by ***, ** and *. 44 Table A4 Alternative Set of Instruments. Dummy Instruments. (1) (2) (3) (4) (5) (6) Panel A: Second Stage Log Average Wage Share non-prod Workers Exports/Sales (EXP ) –0.468 –0.379 –0.399 0.099 0.098 0.096 (0.475) (0.472) (0.475) (0.182) (0.177) (0.183) High Income Exports (HI) 0.32∗∗∗ 0.34∗∗∗ 0.311∗∗∗ 0.123∗∗∗ 0.123∗∗∗ 0.124∗∗∗ (0.107) (0.109) (0.104) (0.044) (0.044) (0.043) Log Sales − − 0.059∗∗∗ − − -0.003 (0.019) (0.007) Panel B1: First Stage (HI) Share BRA exports * 1999 0.227∗∗∗ 0.225∗∗∗ 0.227∗∗∗ (0.044) (0.044) (0.044) Share BRA exports * 2000 0.249∗∗∗ 0.252∗∗∗ 0.249∗∗∗ (0.049) (0.048) (0.049) Share BRA sales * 1999 −0.276∗∗ −0.277∗∗ −0.278∗∗ (0.122) (0.122) (0.122) Share BRA sales * 2000 −0.332∗∗ −0.329∗∗ −0.335∗∗∗ (0.129) (0.129) (0.130) Log Sales − − 0.010 (0.017) R-squared 0.034 0.034 0.034 p-value 0.000 0.000 0.000 Panel B2: Second Stage (EXP ) Share BRA exports * 1999 0.008 0.008 0.008 (0.012) (0.012) (0.012) Share BRA exports * 2000 0.035∗∗ 0.035∗∗ 0.035∗∗ (0.016) (0.016) (0.016) Share BRA sales * 1999 −0.173∗∗ −0.173∗∗ −0.174∗∗ (0.079) (0.079) (0.080) Share BRA sales * 2000 −0.235∗∗ −0.235∗∗ −0.237∗∗ (0.096) (0.096) (0.096) Log Sales − − 0.006 (0.009) R-squared 0.028 0.028 0.029 p-value 0.014 0.007 0.014 Initial Conditions * Year Effects Yes − Yes Yes − Yes Initial Conditions * Exchange Rate − Yes − − Yes − Observations 2544 2544 2544 2544 2544 2544 Number of Firms 901 901 901 901 901 901 Dependent variable in second stage: Log Average Wage (columns 1-3) and Share of non-production workers (columns 4-6). Dependent variables in �rst stage: Exports to high income destinations over Total value of exports (HI), in Panel B1; and Total value of exports over Total value of sales (EXP ) in Panel B2. Alternative instruments: Share of Brazil in Sales interacted with year effects is used instead of the average exchange rate of all trade partners weighted by share in sales. All regressions include �rm and year �xed effects. Bootstrapped SE in parenthesis. Signi�cance at the 1 percent, 5 percent and 10 percent level is denoted by ***, ** and *. 45 Table A5 Alternative Set of Instruments. Exchange Rate Instruments. (1) (2) (3) (4) (5) (6) Panel A: Second Stage Log Average Wage Share non-prod Workers Exports/Sales (EXP ) –0.234 –0.230 –0.141 0.107 0.107 0.101 (0.471) (0.477) (0.468) (0.204) (0.204) (0.203) High Income Exports (HI) 0.323∗∗∗ 0.323∗∗∗ 0.318∗∗∗ 0.128∗∗∗ 0.128∗∗∗ 0.129∗∗∗ (0.113) (0.113) (0.110) (0.044) (0.044) (0.045) Log Sales − − 0.058∗∗∗ − − –0.004 (0.020) (0.007) Panel B1: First Stage (HI) Share BRA Exports * Exchange Rate 0.398∗∗∗ 0.398∗∗∗ 0.398∗∗∗ (0.070) (0.070) (0.070) Share BRA Sales * Exchange Rate −0.500∗∗∗ −0.501∗∗∗ −0.505∗∗∗ (0.190) (0.190) (0.190) Log Sales − − 0.011 (0.017) R-squared 0.032 0.032 0.033 p-value 0.000 0.000 0.000 Panel B2: First Stage (EXP ) Share BRA Exports * Exchange Rate 0.028 0.028 0.028 (0.021) (0.021) (0.021) Share BRA Sales * Exchange Rate −0.330∗∗ −0.330∗∗ −0.333∗∗ (0.140) (0.140) (0.141) Log Sales 0.006 (0.009) R-squared 0.022 0.022 0.024 p-value 0.063 0.063 0.061 Initial Conditions * Year Effects Yes − Yes Yes − Yes Initial Conditions * Exchange Rate − Yes − − Yes − Observations 2544 2544 2544 2544 2544 2544 Firms 901 901 901 901 901 901 Dependent variable in second stage: Log Average Wage (columns 1-3) and Share of non-production workers (columns 4-6). Dependent variables in �rst stage: Exports to high income destinations over Total value of exports (HI), in Panel B1; and Total value of exports over Total value of sales (EXP ) in Panel B2. Alternative instruments: Share of Brazil in Sales interacted with the Brazilian exchange rate is used instead of the average exchange rate of all trade partners weighted by share in sales. All regressions include �rm and year �xed effects. Bootstrapped SE in parenthesis. Signi�cance at the 1 percent, 5 percent and 10 percent level is denoted by ***, ** and *. 46