Global Engagement and the Occupational Structure of Firms

Transcription

1 Global Engagement and the Occupational Structure of Firms Carl Davidson *, Fredrik Heyman +, Steven Matusz *, Fredrik Sjöholm +, and Susan Chun Zhu * March 2016 Abstract: Global engagement can impact firm organization and the occupations firms need. We use a simple task-based model of the firm s choice of occupational inputs to examine how that choice varies with global engagement. We use Swedish data covering all firms and a sample of the labor force for We find a robust relationship, with more globally engaged firms using a labor mix skewed toward high-skill occupations. Moreover, firms are more skill-intensive when they export to distant markets, or when they export differentiated products and capital goods. Using an instrumental variable approach, we find that increased export shares (driven by higher world import demand) skew the labor mix even more toward high-skill occupations. The cross-firm difference in labor mix and a higher share of skilled workers in higher-wage firms contributed to 19% of overall wage dispersion in 1997 and 26% in JEL: F1, F2 Keywords: Occupational structure, Globalization, Multinational Enterprises, Exporters * Michigan State University; + The Research Institute of Industrial Economics, Sweden; Lund University; GEP, University of Nottingham Acknowledgements: We are indebted to Mina Kim, Yi Lu, and Jonathan Vogel for helpful discussions. We have also benefitted from the feedback provided by participants at numerous seminars, workshops, and conferences. Fredrik Heyman and Fredrik Sjöholm gratefully acknowledge financial support from the Swedish Research Council, the Swedish Research Council for Health, Working Life and Welfare, NORFACE, and Torsten Söderbergs Stiftelse. Susan Chun Zhu gratefully acknowledges financial support from the Visiting Fellow Programme at Lund University.

2 1. Introduction Profits are the raison d'être for firms. Toward this end, firms undertake a variety of tasks in addition to production including inter alia research, financial management, logistics, marketing, and sales. Each of these activities, including actual production, requires the input of workers who are trained in various occupations (e.g., production managers, foremen, computing professionals, mechanics, and filing clerks). It is likely that the mix of occupations required to undertake economic activities varies across firms. In particular, the tasks required to support a multinational enterprise (MNE) are likely to differ from those required to support a non- MNE that exports, which in turn are likely to differ from a firm that does not export. However, a decade after Melitz (2003), [t]he productivity of the firm remains largely a black box and we still have relatively little understanding of the separate roles played by production technology, management practice, firm organization and product attributes towards variation in revenues across firms (Melitz and Redding, 2014). In this paper we aim at shedding light on the organization of production within firms, as captured by the distribution of employment across occupations, and its relationship with a firm s international orientation. Our empirical investigation uses comprehensive and detailed Swedish matched employer-employee data spanning 1997 to The data cover all Swedish firms and a representative sample of the labor force, including information on worker occupations at a very detailed level (100 occupations). Merged with the Swedish Foreign Statistics, the data also provide information on firms import and export activities. Initial results are displayed in Figure 1 which shows the aggregate distribution of employment by skill levels for three different firm types: (i) the most integrated ones MNEs; (ii) the least integrated ones Local firms (i.e., non-mnes that do not export); and (iii) the intermediate firm type Exporters (i.e., non-mnes that export). The horizontal axis is the percentile ranking of occupations by skill levels, from the least skilled to the most skilled. The vertical axis is the cumulative employment share of the labor force accounted for by the skill category that is indicated on the horizontal axis. Panel (a) shows occupations ranked according to the average wage for all firms in 1997 and panel (b) is based on a regression ranking derived from Mincer wage regressions (see Section 2.C). Compared to Local firms, the employment distribution for MNEs and Exporters is more skewed towards high-skill occupations (e.g., professionals specialized in finance and sales, computing, and engineering). For instance, Figure 1 reveals that roughly 70% of the employees hired by Local firms are in the 50% lowest ranked occupations. The corresponding figures for MNEs and Exporters are roughly 40 and 50% 1

3 respectively. 1 More elaborate econometric estimations in the paper confirm a robust relationship between the degree of international integration and the distribution of employment at the firm level. The distribution is skewed toward more skilled occupations in MNEs. Non-MNE exporters have a distribution of employment less skewed toward the skilled compared to multinationals, but more skewed toward the skilled compared to Swedish nonexporters. The economic magnitude of this cross-firm difference in labor mix is large. It translates into a 7.8% difference in average wages between MNEs and Local firms, and a 6% difference between Exporters and Local firms. Furthermore, the cross-firm difference in labor mix and a higher share of skilled workers in higher-wage firms, such as MNEs and non-mne exporters, contributed to 19.4% of overall wage dispersion in 1997 and 26% in 2005 (see Section 4.C for detail). Our study also reveals that the labor mix tends to be more skewed toward high-skill occupations when firms mainly export to markets that involve higher fixed costs of export and are more distant from Sweden in terms of geography, or in bilateral trust. In addition, the labor mix is more skewed toward high-skill occupations when firms mainly export differentiated products and capital goods. All these results suggest that a more skilled labor mix may be necessary for a firm to overcome higher costs of entry into a more difficult market. We also investigate whether increased global engagement may have a causal impact on the labor mix. Using an instrumental variable approach, we find that the employment distribution is significantly more skewed toward high-skill occupations when firms have an increased export share driven by higher world import demand. 2 Furthermore, we find a modest shift to a more skilled mix when Local firms switch to be Exporters. Our results are little changed when we (1) control for firm size, productivity, capital intensity, and firm age; (2) control for offshoring and R&D expenditures; or (3) use the average share of college graduates or other alternative methods to rank occupations by skill levels. To help understand the empirical results, we construct a task-based model of the firm s choice of occupational inputs to examine how that choice varies with global engagement. Our model builds on the Chaney 1 Figure 1 is consistent with Matsuyama s model (2007), where production for export is more skill intensive than production for domestic sales since export requires white-collar workers, particularly those with language skills, international business experiences and/or specialists in export finance and maritime insurance. Matsuyama shows that an increase in the world supply of skilled labor increases the degree of globalization. 2 Following Hummels et al. (2014), we instrument for export shares using world-wide shocks to import demand for the relevant product and partner country (see Section 4.E for more detail). 2

4 and Ossa (2013) framework in which production requires a series of tasks to be completed and firms design their production chain to minimize costs. We depart from Chaney and Ossa by assuming that it is more costly to train workers to perform more complex tasks. Within the structure of our model, firms skew employment toward occupations engaged in more complex tasks. Moreover, consistent with our empirical findings, the distribution of employment is more skewed toward high-skill occupations for more globalized firms. Our paper relates to the previous literature that has documented that multinational firms are more skill intensive than domestic firms (e.g., see Markusen 1995 and Barba-Navaretti and Venables 2004 for a survey). Moreover, Bernard and Jensen (1997) and others have provided strong evidence that exporters are more skill intensive than non-exporters. Biscourp and Kramarz (2007), using French firm data, find that an increase in imports results in destruction of production jobs, especially for larger firms. However, due to data limitations, these previous studies usually define production workers as the unskilled and non-production workers as the skilled. Our data allow us to dig deeper into this issue because we have a much finer classification of occupations. We are thereby able to offer new insights that build on previous work. We find that MNEs and Exporters have higher employment shares than Local firms in nearly all high-skill occupations, especially in the category of professionals specialized in finance and sales, computing, and engineering. On the other hand, MNEs and Exporters have lower employment shares than Local firms in almost all low-skill occupations. In addition, our study reveals new patterns of the variation in labor mix across firms that serve different destination markets or specialize in different export products, providing evidence that a higher share of skilled professionals may be necessary for overcoming a higher cost of entry into a difficult market. In addition, using an instrumental variable approach, we document that increased global engagement has a causal effect on the labor mix across firms. Finally, the existing literature largely focuses on manufacturing. In contrast, our study covers the whole economy, including both manufacturing and non-manufacturing industries. This is important since nonmanufacturing industries have been playing an increasingly important role in global businesses. Our paper also relates to a small but growing empirical literature on globalization and firm organization. For instance, Rajan and Wulf (2006) find that U.S. firms have become flatter over time in that they have fewer layers of management. Moreover, Guadulupe and Wulf (2010) find that firms flatten their organizations after trade liberalization by removing layers between the CEO and division managers, and by increasing the number of positions that report directly to the CEO. Caliendo, Monte, and Rossi-Hansberg (2015) use French firm level 3

5 data with information on five different occupation categories three types of management, clerks and bluecollar workers to examine the wage effect of adding a layer (one of the above categories) or by expanding existing layers (see Tåg 2013 for a similar study). New exporters are more likely to add layers than nonexporters, and firms that exit the export market are more likely to drop layers than firms that continue to export. Unlike these empirical studies that focus on organizational hierarchy, our work focuses on the labor mix in firms with different degrees of global engagement. We provide descriptive statistics of the Swedish matched employer-employee data in Section 2 and sketch out a simple theoretical framework in Section 3 to understand the empirical facts revealed by our data. Our model suggests systematic differences in labor mix across firms with different degree of global engagement. We turn to detailed empirical analysis in Section 4 and conclude in Section Data and descriptive statistics We use register-based matched employer-employee data from Statistics Sweden covering the period The firm data contain detailed information on all Swedish firms, including variables such as value added, capital stock (book value), number of employees, wages, ownership status, sales, and industry. Moreover, the Regional Labor Market Statistics (RAMS) provide plant-level information on education and demographics, which we aggregate to the firm level. RAMS include data on all Swedish plants. The worker data cover detailed information on a representative sample of the labor force, including full-time equivalent wages, education, occupation, and gender. 3 Occupations are based on the Swedish Standard Classification of Occupations (SSYK96) which in turn is based on the International Standard Classification of Occupations (ISCO-88). Occupations in ISCO-88 and SSYK96 are grouped based on the similarity of skills required to fulfill the duties of the jobs (Hoffmann, 2004). See Appendix A for details on the occupation classification. Firm level data on export and import by products and countries originate from the Swedish Foreign Trade Statistics, collected by Statistics Sweden. Based on compulsory registration at Swedish Customs, the data 3 The worker data originate from the Swedish annual salary survey (Lönestrukturstatistiken). The survey s sampling units consist of firms included in Statistics Sweden s firm data base (FS). A representative sample of firms is drawn from FS and stratified according to industry affiliation and firm size (number of employees). The sample size consists of between 8,000 and 11,000 firms. The Central Confederation of Private Employers then provides employee information to Statistics Sweden on all its member firms that have (i) at least ten employees and (ii) are included in the sample. Firms with at least 500 employees are examined with probability one. The final sample includes information on around 50% of all employees within the private sector and is representative of the Swedish labor force. See for more details on the data. 4

6 cover all trade transactions from outside the EU. 4 Trade data for EU countries are available for all firms with a yearly import or export of around 1.5 million SEK and above. According to the figures from Statistics Sweden, the data cover around 92% of total goods trade within the EU. Material imports are defined at the 5-digit level according to NACE Rev 1.1 and grouped into Main Industrial Groupings (MIGs) based on intended use. 5 Based on the MIG definition of intermediate inputs we identify offshoring using import data at the firm and product level. Our measure is likely to be more precise than most other measures used in the literature. 6 Information on the nationality of foreign MNEs operating in Sweden comes from the Swedish Agency for Economic and Regional Growth (Tillväxtanalys). The Agency uses definitions of firm nationality that are in accordance with definitions in similar data from the OECD and Eurostat. A firm is classified as a foreignowned MNE if more than 50% of the equity is foreign-owned. Finally, Swedish MNEs are defined as those with international trade in goods or services within the corporation. All data sets are matched by unique identification codes. To make the sample of firms consistent across the time periods, we restrict our analysis to firms with at least 20 employees in the non-agricultural private sector, which are available throughout the period. A. Distribution of firms Firms can be classified by their international integration along different criteria. One possible classification scheme is seen in Figure 2 where the different categories are mutually exclusive. Figures on the share of different firm categories are for the period The first criterion is ownership, where we distinguish between MNEs and non-mnes. MNEs can be Swedish or foreign owned. The next criteria is whether the firm is an exporter or only sells to the domestic market, and the final criteria is whether the firm is engaged in offshoring or not. Most Swedish firms are internationally integrated in at least some respects. We have data on 8,236,835 individual-year observations divided by firm types and industries. About 46 percent of these individuals work in MNEs that both export and offshore and another 24 percent in non-mnes that both export and offshore. In manufacturing industries the majority of employees work in the most internationally 4 Many firms in the non-manufacturing sector export goods as defined in our Customs data. Since the Customs data do not include service trade, firms that only export services are classified as non-exporters. However, as we show in Section 4.G, our results differ little between manufacturing and non-manufacturing industries. 5 MIG is a European Community classification of products. 6 See Hijzen (2005) for an overview of different measures of offshoring. 5

7 integrated firms MNEs that export and offshore, or in non-mne exporting firms that also offshore (not shown). Firms that have low levels of international integration are mainly found in non-manufacturing, especially in Health and Education. Because the vast majority of MNEs are also engaged in both export and offshoring, and non-mne exporters tend to offshore at the same time, we group firms into three types: MNEs defined as multinational firms; Exporters defined as non-mnes that export; and Local firms defined as non-mnes that do not export. We will complement the division by paying special attention to the role of offshoring in Section 4.F. B. How do firms differ in their labor mix? We begin by considering twelve broad occupation categories in terms of their functions in production. Table 1 lists the twelve groups ranked by mean occupational wages. 7 Managers, research/business professionals, and technicians are more skilled (highly educated and better paid) than machine/transportation operators, crafts, clerks, and sales and service workers. Column 2 shows the employment share of these occupation groups. About one-third of total employment is in the occupations that require more skills. To examine how MNEs, non-mne exporters, and Local firms differ in the distribution of employment, we run the following regression (1) λλ jjjjjj = αα kk MM MMMMMM jjjj + αα kk XX EEEEEEEEEEEEEEEE jjjj + ZZ jjjj γγ kk + DD iiii + DD kkkk + εε jjjjjj where ii, jj, kk and t index industries, firms, occupations, and years respectively; λλ jjjjjj is the employment share of occupation k at firm j in year t; MMMMMM jjjj is an indicator of multinational firms (MNE); EEEEEEEEEEEEEEEE jjjj is an indicator of non-mne exporters; ZZ jjjj is a vector of firm characteristics that might affect the employment share, including firm size (the number of employees), capital intensity (capital-labor ratio), labor productivity (value added per worker), and firm age; DD iiii and DD kkkk are occupation-industry and occupation-year fixed effects; and εε jjjjjj is the error term. 8 The occupation-industry fixed effects control for technology differences across industries, and the occupation-year fixed effects control for common macro-level shocks that may affect firm-level employment. 9 7 See Appendix A for more detail about the grouping of occupations into these twelve broad categories. Appendix Table A1 shows which three-digit occupations are included in a specific occupation category. 8 In the rest of the paper our core estimates use employment shares of different occupations. Some empirical work on globalization and labor demand uses changes in wage bill shares as the dependent variable (e.g. Hakkala et al., 2014). We also do so as a robustness test, but results are basically unchanged (see Section 4.G for details). 9 We control for 27 different industries in our occupation-industry fixed effects. 6

8 MM XX Since Local firms are the excluded group, the coefficients αα kk and αα kk represent the difference in employment shares between MNEs and Local firms, and that between non-mne exporters and Local firms, respectively. MNEs and non-mne exporters have relatively high employment shares of high-skill occupations such as managers, research/business professionals, and technicians, as seen in columns 3-4 in Table MNEs and Exporters also tend to have smaller shares of low-skill occupations such as craft, transportation operators, sales and service workers, and laborers. There are, however, two exceptions: MNEs and Exporters have relatively high shares of machine operators and information-processing clerks. This could indicate complementarity between certain low-skill and high-skill occupations. Finally, the difference between MNEs and Local firms is larger than the difference between Exporters and Local firms. C. A more detailed look at the labor mix across firm types Since skill requirements may vary substantially within the 12 broad occupation groups, we turn to occupation categories at the 3-digit ISCO in order to provide a more detailed analysis of the labor mix across firm types. After merging occupations with very few observations, we end up with 100 occupations. We run separate regressions of (1) for each of the 100 occupations. The results are shown in Table A1. To better illustrate the results, we rank occupations by skill levels. Our preferred measure, the beta ranking, is based on estimated Mincer wage regressions at the worker level. For each year we run regressions with individual wages as the dependent variable and with experience, experience squared, gender, education, and occupation at the 3-digit level as independent variables. 11 Industry fixed-effects are also included. We then compute the skill level of an occupation using the estimated coefficient for that occupation plus the estimated coefficient on the education dummy that indicates the median education level of workers in the particular occupation (note that education is measured at 7 levels as described in footnote 11). For example, in Directors and chief executives, the median education level is equal to level 6. Given that the estimated coefficient for this occupation group is 0.92, and the estimated coefficient on the dummy for education level 6 is 0.25, the skill level of Directors and chief executives is Thus, the betas estimated from Mincer wage regressions 10 The one exception is other professionals where the share is relatively high in Local firms. As shown in Table A1, other professional category mainly includes teaching professionals and nursing professionals. 11 Education is measured at 7 levels: (1) Elementary school; (2) Compulsory school; (3) 2 years of upper secondary school; (4) 3 years of upper secondary school; (5) 4 years of upper secondary school; (6) Undergraduate or graduate college education; and (7) Doctoral degree. Education dummies are defined based on the highest level achieved (e.g. Education 6 equals one if the highest level is undergraduate or graduate college education). 7

9 combine the returns to schooling and training required for a specific occupation. In the following analysis we use the ranking for the initial year (1997) of the sample. The results are unchanged when the average ranking over the entire sample period is used. As a robustness check, we use various alternative ways to measure the skill level of occupations, including the share of college graduates and the mean wage for each occupation in Because of the concern about the dominant effect of MNEs on wages, we also compute the mean wage for each occupation using the sample of non-mne firms. 12 Since there are very high correlations between different ranking measures (over 95%), our results are unchanged when alternative measures are used. We rank occupations by skill levels from the lowest (k = 1) to the highest (k = 100). For a particular occupation k = k 0, we compute the cumulative employment share differential between MNEs and Local firms kk 0 MM as MM kk=1 αα kk where αα kk is the difference between MNEs and Local firms in the employment share of occupation MM k estimated from (1). In Figure 3 panel (a) we plot the cumulative employment share differential kk=1 αα kk against the beta ranking of occupation k 0. This curve has two properties. First, since the sum of employment shares across all occupations should be one for any firm type, it is easy to show 100 MM kk=1 αα kk kk 0 = 0, meaning that the curve must meet the horizontal axis at the most skilled occupation ( Directors and chief executives ). Second, the slope of this curve reflects the marginal difference in employment shares (αα kkmm ). 13 If the curve rises at occupation k 0, the marginal difference in employment share of k 0 is positive, meaning that MNEs have a higher employment share of occupation k 0 than Local firms. On the other hand, a declining curve at k 0 indicates a negative αα MM kk0, which implies that MNEs have a smaller employment share of occupation k 0 than Local firms. Note that no restrictions are imposed on the curvature of this curve. Two interesting patterns emerge from this plot. First, the curve declines up to the 60th skill percentile, showing that MNEs have a smaller employment share of these occupations than Local firms. Overall, the employment share of these lower skilled occupations is nearly 25 percentage points lower in MNEs than in Local firms. More than half of the difference is attributable to three occupations: Motor-vehicles drivers (ISCO 12 Some studies find multinational firms to pay higher wages than local firms (see e.g. Heyman et al. 2007). 13 Note that αα MM kk kk0 = 0 MM kk kk=1 αα kk 0 1 MM kk=1 αα kk is the difference between MNEs and Local firms in the employment share of occupation kk 0. This is the marginal difference in employment shares. Thus, if αα MM kk0 < 0 meaning that MNEs have a smaller employment share of occupation kk 0 than Local firms, the cumulative employment share differential declines i.e., kk 0 MM kk kk=1 αα kk < 0 1 MM kk=1 αα kk. Similarly, the cumulative employment share differential increases if MNEs have a bigger employment share of a particular occupation than Local firms. 8

10 832; αα MM kk = 5.95%); Building finishers and related trade workers (ISCO 713; αα MM kk = 4.46%); and Shop, stall and market salespersons and Fashion and other models (ISCO 521, 522; αα MM kk = 5.39%). On the other hand, this lower-end portion of the curve also has a small jump at Material-recording and transport clerks (ISCO 413; αα MM kk = 2.01%). This difference in employment share is relatively large compared to the average employment share of 1.96 percent for the whole economy, and could arise from the need by MNEs to operate a more complex production and sales network than Local firms. Second, for the th skill percentile, the curve is almost monotonically increasing, which indicates that for these occupations, MNEs have higher employment shares than Local firms (see footnote 13). Since these higher skilled occupations are mostly professionals and managers, the result suggests that the pattern displayed by Table 1 holds for more detailed occupation classifications. As shown in Table A1, the employment share difference is largest for professionals specialized in finance, sales and business (ISCO 341, 241), computing (ISCO 213, 312), and engineering (ISCO 311, 214). The difference in the employment share for these six professional occupations is almost 16 percentage points between MNEs and Local firms. Figure 3 panel (b) displays the cumulative employment share differential between non-mne exporters and Local firms against the beta ranking of occupations. The key patterns are similar to those in panel (a). Compared to Local firms, non-mne exporters have smaller employment shares of less skilled occupations and higher employment shares of more skilled occupations. Comparing the two plots in panels (a) and (b), we can see that MNEs have a distribution of employment even more skewed toward skilled, and the difference from Local firms in skill distribution is even larger. MM XX In panels (c) and (d) we show the plots based on the estimates of αα kk and αα kk when firm characteristics are controlled for. After adding the controls, the difference in skill distribution between MNEs and Local firms becomes slightly smaller, suggesting that part of the difference between MNEs and local firms is due to differences between these firm types in terms of firm productivity, size, capital intensity, and firm age. However, the key patterns remain. A similar observation can be made for panel (d) for the difference between non-mne exporters and Local firms. 3. A Conceptual Framework Before continuing our exploration of the data, we develop a theoretical framework to better understand 9

11 why we might expect systematic differences in the occupational structure of employment across firms with different characteristics. 14 Our model builds on the insights of Chaney and Ossa (2014) in which firms design their production chain to minimize costs. The idea behind the Chaney and Ossa framework is simple. Producing output requires that a series of tasks be completed. Firms create teams of workers and assign them a set of tasks to complete. In doing so, the firm trains workers to have a team-specific core competency or specialization, which allows them to complete a certain task at zero cost. Team members complete related tasks that are close to their specialization at positive cost, with the cost increasing in the distance of that task from the team s core competency. Thus, we can think of a firm hiring a group of attorneys and training them in the law that relates specifically to the firm s product market. Those attorneys could also deal with other legal issues within the firm that are somewhat related to their specialization. The firm s goal is to select the number of teams and assign them tasks in a way to produce the desired output as cheaply as possible. In Chaney and Ossa (2014) all tasks are symmetric. Our first point of departure is to assume that tasks can be ranked in order of complexity and that it is more costly to train workers to perform more complex tasks. We could model the increase in costs in a variety of ways. For example, we could assume that firms pay higher wages to compensate workers for the effort entailed in mastering more complex tasks, or we could assume that more complex tasks require additional complementary resources in the training process. Perhaps the simplest assumption is that mastering more complex tasks utilizes more efficiency units of labor. It is this latter assumption that we adopt here. Our second deviation from Chaney and Ossa is just a matter of interpretation. To operationalize the theory, we map teams to occupations by defining an occupation as a specialization combined with the tasks assigned to the team. For example, firms hire teams of attorneys, accountants, mid-level managers, machinists, and so on. This interpretation is consistent with the definition of occupations in the context of ISCO-88 and SSYK96 where occupations are grouped together and aggregated on the basis of the similarity of skills required to fulfill the tasks and duties of particular jobs (see Hoffman, 2004). Given our interpretation, we can then use the terms occupation and teams interchangeably. Formally, we assume that production requires the completion of a set of tasks ωω [0,1] where we order 14 See also Davidson et al. (2015) for a more elaborated theoretical model. 10

12 tasks in terms of complexity so that ωω is more complex than ωω for all ωω > ωω. Each task must be completed once in order to produce one unit of output. As discussed above, we assume that a fixed amount of labor is required to master each task and this amount is increasing in the complexity of the task. Workers employed in an occupation undertake a range of tasks in addition to their core competency, but the cost of doing so is increasing in distance from their specialization. Specifically, suppose that there are nn occupations indexed by the kk, let ss kk [0,1] represent the specialization associated with occupation (and team) kk and let l kk represent the number of workers in occupation kk needed to produce the set of tasks ωω ωω kk, ωω kk required to produce yy units of output. We assume that ωω (2) l kk = ff + ββss kk + yy kk ss kk ωω dddd, ff > 0, ββ > 0. ωω kk In (2), the first two terms capture the fixed cost of establishing and training the team. The last term represents the labor hired to complete production of y units of output. Our assumption that ββ > 0 is the only substantive departure from Chaney and Ossa. As we show below, this assumption results in systematic differences in the occupational structures between firms with different characteristics. Finally, for simplicity, we assume that all workers earn the same wage (normalized at 1). The firm s objective is to design its production chain to minimize the cost of producing y units of output. In other words, it must decide how many teams to form (nn), the location of each team s specialization (ss kk ) on the task continuum, and the set of tasks to assign each team (ωω kk and ωω kk ). Formally: nn ωω kk (3) minimize ff + ββss kk + yy nn, ss kk, ωω kk, ωω kk ss kk ωω dddd. kk=1 ωω kk The design of the production chain determines the occupational mix of workers hired. Our goal is to examine how that mix changes with global engagement. Following Helpman, Melitz and Yeaple (2004), we assume that increased global engagement is associated with firms that are bigger and that have greater fixed costs of production and market entry when compared with their domestic counterparts. We also assume that the additional fixed costs are skewed toward more complex tasks. For example, additional fixed costs associated with exporting might be concentrated in tasks dealing with writing contracts and managing exchange risk rather 11

13 than product assembly. Thus, increased global engagement leads to increases in both yy and ββ. 15 Increases in y capture the size effect, while changes in ββ capture the direct effect of greater global engagement, holding size fixed. Hence, our model allows us to separate the size effect from the pure global-engagement effect. Solving the firm s problem is best done via backwards induction in three steps. First, for each team, solve for the optimal specialization (ss kk ) given that team s assigned set of tasks (defined by ωω kk and ωω kk ). Second, given the number of teams (n), solve for the optimal set of tasks to assign to each team. Finally, solve for the optimal number of teams. We provide a detailed analysis of the model, its solution and the impact of global engagement in our companion paper Davidson et al. (2015). Here, we provide a summary of the main results and the driving forces behind them. The first insight that we obtain is that for a given set of tasks, the firm will select a specialization closer to the least complex task than it is to the most complex task in particular, it can be shown that ss kk = 1 ωω 2 kk + ωω kk ββ < 1 ωω yy 2 kk + ωω kk. 16 The intuition for this result, which we refer to as the specialization bias, is as follows. With a uniform wage across workers, the firm s goal is to minimize employment. If ββ = 0 so that the fixed component of employment is independent of the team s specialization, as in Chaney and Ossa (2013), the firm would choose ss kk at the mid-point of the range of tasks in order to minimize the variable employment needed to complete the tasks. However, this is not the case if the fixed labor requirement increases with complexity. In this case, the firm can save on fixed employment by marginally reducing the specialization that it uses, but the cost of doing so is an increase in the variable component of employment. A given reduction in ss kk results in a larger reduction in the fixed component of employment relative to the increase in variable employment when ββ is large or yy is small. Thus, an increase in ββ (the direct effect) due to increased global engagement, makes the specialization bias more pronounced, while an increase in y (the size effect) moderates the specialization bias by increasing the relative importance of variable costs. Much of what follows hinges on the tradeoff between fixed and variable costs and how that tradeoff is affected by global engagement. The second insight concerns the range of tasks assigned to each team. Since the fixed cost of 15 Suppose we were to merge our framework with a Melitz (2003)-style model. All firms with a given degree of globalization (for example, all exporters) would have the same value of ββ, which would be larger than the ββ for lessglobalized firms. However, more productive exporters would be larger than less productive exporters. 16 Formally, ss kk minimizes (3) given ωω kk and ωω kk. See Davidson et al. (2015) for details for parameter restrictions that ensure an interior solution to this problem. 12

14 establishing a team is increasing in the complexity of that team s specialization, the firm will assign teams with less complex specializations a narrower range of tasks than those with more complex specializations. To show this, we first note that since each task must be completed, the most complex task assigned to the team with specialization ss kk must be the same as the least complex task assigned to the team with specialization ss kk+1 (i.e., ωω kk = ωω kk+1 ). With this in mind, if we substitute ss kk into (3), fix n, and optimize over ωω kk and ωω kk we can show that ωω kk = ωω kk+1 = kk kk(nn kk) ββ. We can now solve for the range of tasks assigned to team k: ωω nn yy kk ωω kk = 1 [nn 2kk + 1] ββ, which is increasing in k. nn yy 2kk 1 2nn We can push the analysis further by substituting the optimal cut-off values back into ss kk to obtain ss kk = nn + (nn kk)(kk 1) ββ, implying that ss 2 yy kk+1 ss kk = 1 [nn 2kk + 2] ββ, which is increasing in k that nn yy is, the distance between specializations is increasing in the complexity of the task, so that the design of the production chain is skewed towards less complex occupations. In our companion paper, we use the Groeneveld and Meeden coefficient to measure the degree of skewness in the distribution of specializations and show that ββ it is approximately equal to 1 3 yy nn2 ; so that the skewness is positively related to the pure globalization effect and inversely related to the size effect. 17 As ββ increases, the relative cost of using more complex occupations rises, leading firms to use relatively more occupations with less complex specializations and assigning them a narrower range of tasks, therefore increasing the skewness of the distribution of specializations. 18 The size effect works in the opposite direction, since any increase in y makes variable costs relatively more important and decreases the firm s incentive to shade each team s specialization towards low-skilled tasks. To tie the model s results back to our earlier empirical results, we first note that our empirical work uses the language of skill when referring to tasks in contrast to the language of complexity in the theoretical model. We think that it is a small step to imagine that undertaking more complex tasks entails greater skill. Indeed, we can attribute the additional fixed cost for training teams to undertake more complex tasks as being 17 The Groeneveld and Meeden Coefficient is defined as SSSS (μμ νν)/ee ss kk νν 2, where μμ and νν are the mean and the median of the occupation distribution, respectively. We use this measure of skewness due to its tractability. Other measures of skewness produce qualitatively identical results. 18 The specialization bias also increases. We show in Davidson et al. (2015) that the increased specialization bias is more pronounced for teams specialized in less complex tasks. This effect reinforces the impact of a change in ββon the skewness of the distribution of specializations. 13

15 due to extra time or resources necessary to become more skilled. Having said that, we note that the fact that the model predicts that the distribution of specializations is skewed in favor of low-skilled occupations is not directly related to Figure 1, since that figure shows the cumulative distribution of employment across occupations for different types of firms. However, given our framework, the skewness of the specialization distribution does have strong implications for employment patterns across skill levels. In particular, since lessskilled workers are assigned tasks that are near their specialization, the amount of variable employment involved is minimal. In contrast, more-skilled workers are assigned a wide range of tasks, some of which are quite distant from their specialization. This entails the need for a substantial amount of variable employment. Combined with the assumption that more complex occupations entail more fixed employment, we conclude that employment in more complex occupations is larger than employment in less complex occupations. In other words, employment will be skewed towards high-skilled occupations. Formally, we can derive the theoretical analog of Figure 1. If we substitute the optimal cut-off values and specializations back into (2), we can derive the number of workers employed in occupation k. We obtain l kk = ff + yy 4 1 nn 2 + ββ2 +4nn+2 yy nn2 4 occupation index to obtain: ββ(nn+2) 2nn + 2ββ nn 2ββ2 (1+nn) yy kk + 2ββ2 kk2. Differentiate with respect to the yy (4) l kk = 2ββ nn 1 ββnn2 yy 2ββ2 + (2kk 1). yy In our companion paper, we show that nn 2 < yy ββ is the second-order condition for the firm s optimization problem. Therefore, ββ > 0 implies that l kk > 0 and 2 l kk > 0 for all occupations kk 1. Occupation-specific kk 2 employment is increasing and convex in the occupation index. Moreover, since the share of occupation-specific employment is simply occupation-specific employment divided by the firm s total employment, convexity of occupation-specific employment directly implies that the cumulative distribution function for occupation shares is also convex. In addition, the Arrow-Prat measure of convexity is increasing in ββ and decreasing in yy. 19 To summarize, the pure globalization and size effects influence the skewness in employment through their impact on the distribution of specializations. It follows that the skewness in employment should be 19 This is seen by using (4) to solve for 2 l kk kk 2 l kk. After some simplification, 2 l kk = 14 kk 2 l kk 2nn. yy ββ nn 2 nn+2nnnn

16 positively related to the pure globalization effect and inversely related to the size effect. Moreover, the cumulative distribution of employment will be more convex as employment becomes more skewed toward more complex specializations. We can address these predictions in our empirical work by explicitly including firm size and degree of global engagement as explanatory variables. The last step in solving the firm s problem is to substitute the optimal values for the specializations and the cut-offs back into (3) and minimize over n, the number of teams (and occupations). In our companion paper, we show that, as one would expect, firms generally create more teams as they grow in size. Intuitively, larger firms should employ a greater number of occupations since an increase in output makes minimization of variable employment a greater priority. Similarly, an increase in ββ raises the cost of adding additional occupations, leading the firm to use fewer occupations. The key question is whether our results linking the pure globalization effect and the size effect to the employment distribution across occupations continue to hold when the number of teams is endogenous. The answer is not straightforward because, as we pointed out above, the degree of skewness of the specialization distribution is linked directly to ββ yy nn2. Thus, changes in ββ and y have a direct effect as well as an indirect effect on the degree of skewness, with the latter arising via their influence on the optimal number of occupations. Given that n is generally decreasing in ββ and increasing in y, these two effects work in opposite directions. Fortunately, in our companion paper Davidson et al. (2015) we are able to show that the direct effect always dominates both for changes in ββ and y. 20 Thus, all of our qualitative results generalize to the case in which n is chosen optimally. In our empirical work, we have direct measures of firm size, but we do not have direct measures of ββ. Rather, we have measures that capture the degree to which a firm is globally engaged (e.g., the indicators for MNEs and Exporters and also export shares), and attributes of the market that a firm mainly serves (e.g., distance of foreign markets from Sweden). We conjecture that relatively more workers specialized in complex tasks are needed as firms are more globally engaged, and try to penetrate more distant markets. In each case, we would expect ββ to be larger and the distribution of employment within the firm to be more skewed toward the high skill occupations. We explore these conjectures in the remainder of the paper. 20 An application of the Envelop Theorem allows us to place an upper bound on the indirect effects. 15

17 4. Firm-level Results A. Empirical specification The primary goal of our empirical analysis is to understand how the labor mix varies across firms with different degree of global engagement. To capture the complexity of tasks undertaken by workers in different occupations, we measure the skills embodied in different occupations. As described in Section 2.C., we measure the skill level of occupations, indexed by k, in three ways: (1) the average log wage for k; (2) the average share of college graduates for k; and (3) the skill percentile ranking of k (a higher k meaning a more skilled occupation), where occupations are ranked based on the betas estimated from Mincer wage regressions, average wages of all firms, or average wages of only non-mnes. We then construct an index to summarize the skill level of the labor mix at a firm. 21 Specifically, we compute SS jjjj, the skill index for firm j in year t as (5) SS jjjj = kk λλ jjjjjj ss kk where ss kk represents the skill level of occupation k and λλ jjjjjj is the employment share of occupation k at firm j in year t. The skill index is higher if employment is allocated more toward higher skilled occupations. As discussed in Section 2.C, compared to Local firms, MNEs and Exporters have a smaller employment share for almost all low-skill occupations (in the 0-60% skill percentile) while having a higher employment share for all high-skill occupations (in the % skill percentile). Thus, although our skill index is computed as the average skill level of employment, a higher index for MNEs and Exporters in fact reflects a shift in the distribution of employment toward more skilled occupations, which is also clearly displayed in Figure 1. Moreover, since ss kk is fixed for a specific occupation k, the cross-firm difference in the skill index reflects the difference in labor mix rather than the skill difference within occupations. Finally, the index based on the skill percentile ranking 21 An alternative approach is to impute the task content of occupations (see e.g. the seminal work by Autor et al and Acemoglu and Autor 2011 for an overview). Two studies that look at how globalization affects the relative demand for different job tasks are Becker et al. (2013) (analyzing the impact of offshoring) and Hakkala et al. (2014) (studying multinational activity and inward FDI). As pointed out by Acemoglu and Autor (2011), the intensity of use of non-routine cognitive ( abstract ) tasks is highest in professional, technical and managerial occupations, and lowest in service and laborer occupations. Routine cognitive tasks are most intensively used in clerical/sales occupations and routine manual tasks are most prevalent in production/operative occupations. Non-routine manual tasks are most intensively used in production, operative, and service positions. Based on their characterization of the major task differences across broad occupational groups, our results in Table 1 and Figure 3 suggest that compared to Local firms, MNEs/Exporters have a higher employment share of high-skill occupations that are intensive in non-routine cognitive tasks while having a lower employment share of almost all lower skilled occupations that range from being intensive in routine cognitive/manual tasks to being intensive in non-routine manual tasks. Therefore, our results suggest that the skill content of occupations better characterizes the difference in labor mix between MNEs/Exporters and Local firms in our sample, and hence in the following analysis we focus on the skills embodied in different occupations. 16

18 of occupations is bounded between zero and one, and similar to the index used by Zhu and Trefler (2005) to measure the skill content of a country s exports. A value of 0.5 indicates that employment is evenly distributed across all occupations. 22 The main specification is (6) SS jjjj = δδ MM MMMMMM jjjj + δδ XX EEEEEEEEEEEEEEEE jjjj + ZZ jjjj γγ + DD ii + DD tt + μμ jjjj, where DD ii and DD tt are industry and year fixed effects respectively; and μμ jjjj is the error term. Local firms are the excluded group. Thus, δδ MM represents the difference in the skill index between MNEs and Local firms, and δδ XX represents the skill difference between non-mne exporters and Local firms. We note that the coefficients estimated using equation (6) are closely related to those obtained from the regressions separately for 100 occupations as specified in equation (1). Specifically, given the definition of SS jjjj, it is straightforward to show that δδ MM and δδ XX MM XX are weighted sums of αα kk and αα kk respectively, where the weight is ss kk (the skill level of occupation k). B. Baseline Results Table 2 displays the estimation results. In column 1 the skill index is computed using average log wages for 1997 as a proxy for skills. The estimated coefficients on MNE and Exporter indicate that relative to Local firms, the average wage is about 7.8% higher by MNEs and 6% higher by Exporters. Note that these wage differentials reflect the difference in occupational structure across firm types rather than the wage gap between MNEs/Exporters and Local firms within the same occupation. Since employment at MNEs and Exporters is skewed toward more skilled and better paid occupations, these numbers show us on average how much more MNEs or non-mne exporters need to pay their workers given their difference in the skill distribution of employment from Local firms. There is a concern that wages could be confounded by rent sharing between firms and workers (e.g. Frías et al and Helpman et al. 2012). Thus, in column 2 we measure the skill level of occupation k using the average share of college graduates for k in The coefficients on MNE and Exporter suggest that compared to Local firms, the share of college graduates is on average 4 percentage points higher in MNEs and 22 This is a limiting result as the number of occupations tends to infinity. For a finite number of occupations, the lower limit of the index is 1 KK, and SS jjjj = 1 KK+1 when λλ 2 KK jjjjjj = 1 for all k. KK 17