International Trade and Job Polarization: Evidence at the Worker Level Wolfgang Keller1 and Hˆale Utar∗2 1
University of Colorado and NBER 2
Bielefeld University July 15, 2015
Job polarization is the shift of employment and earnings from mid-level wage jobs to both high- and lowwage jobs. Using longitudinal employee-employer matched data covering all residents and workplaces in Denmark, we study the effect of international trade on job polarization as trade barriers fell away with China’s entry into the World Trade Organization. We show that trade can explain the U-shaped pattern of employment changes that is characteristic for job polarization. For the large mid-level wage group of machine operators, already adversely by automatization and the introduction of robots, import competition leads to a loss of eight months of mid-wage employment, and increases of both low-wage and high-wage employment of about four months and one month respectively over a period of eight years. Trade leads to job polarization mainly by shifting workers from initially abundant manufacturing jobs to both high- and low-paying services jobs. Trade leads to less job polarization for women than for men even though women experience overall job polarization more than men, a finding that we relate to the swiftness of the reduction in labor demand caused by import competition. Furthermore, trade has increased gender inequality by shifting women to a lesser extent into high-wage jobs than men. We discuss a number of reasons that might be behind these findings, as well as possible policy implications. ∗
The study is sponsored by the Labor Market Dynamics and Growth Center (LMDG) at the University of Aarhus. Support of the Dept. of Economics and Business, Aarhus University and Statistics Denmark are acknowledged with appreciation. We thank Henning Bunzel for facilitating the access to the confidential database of Statistics Denmark and for his support, David Autor for his insightful discussion of the paper and Nick Bloom, Susanto Basu, Dave Donaldson for helpful comments and suggestions.
1
Introduction
The recent transformation of China from a rather poor and largely closed economy to one of the largest manufacturing producers over little more than two decades has been among the most significant changes in the world economy. China’s share of world manufactures exports rose from 2 % to 4 % over the years 1990 to 1999, only to accelerate pace and account for 16 % of world exports in manufacturing goods by the year 2013.1 For manufacturing producers in rich countries such as Denmark, the emergence of China is felt as a major competitive shock. This is particularly true in the labor-intensive textiles and clothing sector, where the traditional comparative advantage of low-wage countries has been compounded by trade liberalization in form of import quota removals through China’s entry into the World Trade Organization (WTO) in the year 2001. In this paper we study the employment trajectories of Danish workers to determine whether import competition from China has caused job polarization. Job polarization is the shift of employment shares from mid-level wage jobs to both high- and low-wage jobs. This phenomenon is among the most significant labor market developments in recent decades because it has been observed in many rich countries.2 Figure 1 shows that while job polarization has been felt strongly in the United States, developments in Denmark have been even more dramatic. Mid-level employment share losses affect 60 % of all occupations (40 % in the U.S.), and it occurred at a faster pace (seven more years are shown for the U.S.).3 This means that Denmark may provide important new lessons on job polarization. Our focus will be on the textiles and clothing industry, which faces 1
Authors’ calculation using World Development Indicators dataset of the World Bank. For the Unites States, see Autor, Katz, and Kearney (2006, 2008), Autor and Dorn (2013); United Kingdom: Goos and Manning (2007); Germany: Spitz-Oener (2006), Dustmann, Ludsteck, and Schonberg (2009); France: Harrigan, Reshef, and Toubal (2015) and across 16 European countries, see Goos, Manning, and Salomons (2014). 3 The figure for Denmark is based on all employees in Denmark aggregated by 3-digit occupation level (International Standard Classification of Occupations) excluding agriculture. The figure for the United States is taken from Autor and Dorn (2013) and it covers all nonfarm employment in the US. Initial occupational mean wage is based on hourly wage data from 1991 for Denmark and for 1980 for the U.S. 2
2
major import competition from China. In our analysis the effect of import competition turns on whether a specific worker is employed by a firm manufacturing a product for which China’s entry into the WTO removes binding import quotas. We study workers’ employment (and to a lesser extent earnings) trajectories over the years 2002 to 2009. Covering the relatively high rates of export growth during China’s first eight years of WTO membership, this period captures the short and medium term implications of increased import competition. The end of our period of observation is determined by data considerations.4
-.2
100 x Change in Employment Share .6 0 .2 .4
Smoothed Changes in Employment by Wage Percentile
0
20 40 60 80 Wage Percentile (Ranked by Initial Occupational Mean Wage)
100
1980-2005 U.S. 1991-2009 Denmark
Figure 1: Smoothed Changes in Employment in Denmark and in the United States
We employ administrative longitudinal data on the universe of persons aged 15 to 70 years old in Denmark to study the impact of import competition on worker’s movements between jobs.5 The data set is unusually rich in providing information for each individ4
Occupational categories were changed in the year 2009. The Great Recession starting in 2008 took place too late to be driving our results. 5 Our data are not the administrative records themselves but an integrated database created by Statistics Denmark from administrative registers. They are owned by public authorities and supple-
3
ual worker on his or her employer, hours worked, unemployment history, and level of education, among other variables. Information on the tasks workers perform is available at a relatively detailed four-digit level (International Standard Classification of Occupations, ISCO). We can therefore distinguish workers not only by industry and education but also by the tasks they perform. Because we want to exploit a specific trade policy shock–the expiration of the Multi-fiber Arrangement (MFA) quotas as China acceded the WTO–we focus on workers who initially were employed in Denmarks textiles and clothing sector.6 The original purpose of the MFA in the year 1974 was to provide comprehensive protection against competition from low-wage country exports of textiles and clothing through quantitative restrictions. After years of multilateral negotiations, it was agreed in the year 1995 that the MFA would gradually be lifted. China’s non-WTO status rendered it ineligible to benefit from these trade liberalizations, which changed only once China had joined the WTO in the year 2001. The subsequent dramatic surge of Chinese textiles and clothing exports and the resulting increase in the competition is the plausibly exogenous source of shifts in employment trajectories among Danish workers.7 An advantage of studying the incidence of trade in a single sector is that it eliminates the influence of demand shocks, technology shocks, or secular trends that might be correlated with imports in a multi-sector analysis. Our focus on workers with initial employment in the textiles and clothing sector comes at the cost of a smaller sample than we would have studying the entire economy. At the same time, we will see that not only are the forces that drive job polarization in Denmark’s overall economy present in mented with additional information drawn from various surveys (see Bunzel 2008, Timmermans 2010). Earlier work using this data set includes Groes, Kircher, and Manovskii (2015), Hummels, Jorgenson, Munch, and Xiang (2014). 6 The increase in trade with China associated with the MFA quota removals has been employed recently to study technology upgrading of European firms (Bloom, Draca, and van Reenen 2014), as well as in Utar (2014, 2015) to study firms’ and workers’ adjustment. These studies do not focus on job polarization. 7 Utar (2014) shows that the quotas were binding, and that their removal led to increased import competition. Brambilla, Khandelwal, and Schott (2008) examine the effects of the quota lifting in the U.S.
4
the textiles and clothing industry, but movements outside of textiles and clothing (and manufacturing) play a key role in trade’s effect on polarization. We contribute to a growing literature on the importance of international trade in explaining job polarization, much of which examines changes at the industry or region level. Recent work on local labor markets has shown that international trade and offshoring have not played a major role for job polarization in the U.S. (Autor and Dorn 2013), a result that has been confirmed in other rich countries (Goos, Manning, and Salomons 2014, Michaels, Natraj, and van Reenen 2014). Job polarization is mainly explained by information and communications technology investments whereby computers, robots and automatization replace workers that perform easily programmable tasks (so-called routine-biased technical change; Autor, Katz, Kearney 2006, Goos and Manning 2007).8 Our work advances the literature by shifting the focus from aggregate labor market consequences to job changes at the worker level. By examining the difference in responses between workers who are ex ante observationally similar except that one works for a firm that will compete with Chinese imports while the other worker does not, we provide causal evidence on the impact of trade on employment polarization. By shedding new light both on the nature of frictions that workers face when they move between jobs (such as loss of occupation-specific human capital), this paper contributes to a better understanding of worker welfare and the overall welfare effect of international trade. We provide one of the first accounts of job polarization using micro level data for any country. The case of Denmark illustrates well the widely-shared experience in rich countries from a world characterized by skill-biased technical change to one characterized by job polarization. The paper shows that both worker’s education and gender were important in shaping the emergence of job polarization in Denmark. This paper is in a long tradition of work examining the role of international trade for 8
An exception is Firpo, Fontin, and Lemieux (2011) who find offshoring to be important. Other studies of job polarization that do not specifically address the role of international trade include SpitzOener (2006) and Dustmann, Ludsteck, and Schonberg (2009).
5
labor market outcomes, including the increase in relative wage of educated workers in many countries during the 1970s and 1980s (see Feenstra 2000). The focus in recent open economy analysis has shifted to tasks instead of industries or education levels (Grossman and Rossi-Hansberg 2008), which is line with evidence showing that worker outcomes have become strongly related to occupation characteristics (Acemoglu and Autor 2010). By affecting the relative labor demand for mid-level versus high- and low-wage jobs, international trade could in principle be a cause of job polarization. This would require that the factor content of imports from low-wage countries is intensive in tasks for which mid-level wages are paid in rich countries. At the same time it can be challenging to separate technology from trade explanations for job polarization, especially when trade induces technical change (Bloom, Draca, and Van Reenen 2014, Utar 2014). We seek to make progress on the causal role of trade by using a specific trade policy shock together with worker-level information on shifts between occupations and sectors. Our main finding is that international trade plays a statistically and economically significant role in explaining job polarization. Imports from China reduced the employment of a machine operator, a typical mid-level wage textile worker, by about eight months over the period 2002 to 2009. Trade with China also increased employment of machine operators in both high- and low-wage jobs by one month and four months respectively. Trade can explain the U-shaped employment pattern in Figure 1. While this contrasts with earlier work on the causes of job polarization, it is in line with recent work documenting a powerful effect of the China trade for labor market outcomes in rich countries.9 In particular, our finding reconciles worker-level evidence on the sizable effect of import competition from China (Autor, Dorn, Hanson, and Song 2014, Utar 2015) with the scant evidence for trade causing job polarization in studies using more aggregate data.10 Trade leads to job polarization mainly by pushing workers from initially abundant mid9
See Autor, Dorn, and Hanson (2013), Bernard, Jensen, and Schott (2006), Pierce and Schott (2013), Utar (2014); also Utar and Torres-Ruiz (2013), and Hummels, Jorgenson, Munch, and Xiang (2014). 10 Autor, Dorn, and Hanson (2014) present evidence that technology and trade both affect outcomes at the level of US local labor markets; their study does not focus on job polarization.
6
level jobs in manufacturing towards both high- and low-wage jobs in services. We evaluate the effect of trade alongside technology and offshoring explanations of job polarization, finding that the trade effect is robust and hardly affected quantitatively by other explanations. Trade affects most strongly manual labor, workers performing both routine and non-routine tasks. The paper also provides a number of interesting results on gender differences in the experience of job polarization. We find that while women experience overall job polarization more than men, job polarization caused by trade affects women less than men. Key to understanding this, we believe, is that trade has affected the labor market in a more concentrated way than technical change did. First, trade led to a reduction of labor demand mostly in manufacturing, whereas technical change has changed labor demand both in manufacturing and non-manufacturing. Second, while technical change has gradually changed the workplace in industrialized countries since the 1980s, the impact of trade was felt more suddenly around China’s entry into the WTO. The higher costs of transitioning for women in the presence of a sudden drop in labor demand may be due to differences in the availability of re-training possibilities, or because women have on average a lower labor market attachment than men. The next section discusses the facts on job polarization in Denmark during the 1990s and 2000s. Section 3 introduces a number of possible causes of job polarization by presenting a number of micro facts on individual worker transitions between occupations in Denmark’s textiles and clothing sectors. Section 4 provides background on Denmark’s textiles trade as China entered the WTO, and it describes the data that will be employed. Regression results on the role of trade for job polarization are presented in section 5. In this section we also consider technical change and offshoring as alternative explanations of job polarization, and examine the extent to which trade has affected men and women differently. Section 6 provides a concluding discussion. An number of additional results are shown in the Supplementary Appendix.
7
2
Job Polarization in Denmark
This section describes the overall patterns of labor demand in Denmark that can be distinguished over the years 1991 and 2009 to set the stage for our analysis. We then examine demand shifts between broad sectors before focusing on Denmark’s textiles and clothing workers in the years 2001 and 2009.
2.1
Major patterns of Denmark’s labor demand since the early 1990s
Over the last decades the nature of labor demand has changed in industrialized countries, and Denmark is no exception to this. This is made strikingly clear by considering the decade before and after the year 2000 separately. Figure 2 presents the smoothed changes in employment shares ranked by mean initial occupational wage for all of Danish workers outside agriculture, as in Figure 1, except that we distinguish the years 1991 to 2000 from the period 2001 to 2009. During the 1990s, we see that labor demand was increasing with wages. To the extent that we can think of wage as a proxy for worker skills, this pattern implies that skilled workers have tended to benefit while unskilled workers have fallen behind during the 1990s. It has been shown that skill-biased technical change was the most important driver of this pattern (Acemoˇglu 2002). While a lower relative demand for less skilled workers could be important in raising inequality, the policy implication from the 1990s is clear: if skills can shield a worker from declining labor demand, investing in skills, particularly through education, is a plausible welfareenhancing policy. The pattern for the years 2000-2009 in Figure 2 shows that Denmark’s workers now live in a different world. Growth of jobs is increasingly concentrated at the tails of the occupational wage distribution, while the middle of the distribution is hollowed out. The employment growth pattern in Denmark is U-shaped in the 2000s, as in 8
many industrialized countries. We are interested in better understanding the reasons for this, and in particular, the role played by international trade. That would also provide relevant information for policy makers, at least with trade as an additional dimension that should be considered in the formulation of labor market policies.11
-.4
100*Change in Employment Share -.2 0 .2
.4
Smoothed Changes in Employment by Wage Percentile
0
20 40 60 80 Wage Percentile (Ranked by 1991 Occupational Mean Wage) 1991-2000
100
2000-2009
Figure 2: Smoothed Changes in Employment in Denmark by Occupational Wage Percentile
2.2
Shifts between manufacturing and services, 2000-2009
Given this U-shaped pattern that is the hallmark of job polarization, we follow the literature and distill the pattern into changes for three separate groups, called low-, mid-level, and high-wage workers (Autor 2010, Goos, Manning, and Salomons 2014). The groups are formed based on the median wage paid in an occupation in an initial 11
While this paper does not focus on policy, we touch on some of the issues in the conclusions. Autor (2010) discusses economic policy issues in the light of job polarization in the U.S.
9
year.12 The high wage occupations comprise of managerial, professional, and technical occupations. Mid-level wage occupations are clerks, craft and related trade workers, and plant and machine operators and assemblers. Finally, low-wage occupations include service workers, shop and market sales workers, as well as elementary occupations.13 These wage categories capture also broad differences in terms of skills across workers, although each wage group encompasses workers in quite different occupations performing rather diverse sets of tasks.
-.2
-.1
0
.1
.2
Changes in Occupational Employment Share, 2000-2009
Low-Wage
Mid-Wage
High-Wage
Manufacturing Workers in 1999 Service Sector Workers in 1999
Figure 3: Changes in employment shares of low-, mid-, and high wage occupations by workers’ initial industry, 2000-2009
Figure 3 analyzes employment growth for these three wage groups in Denmark’s manufacturing and services sector during the years 2000 to 2009 (at the worker-level).14 12
We rank major ISCO occupations using the median log hourly wage in 1991 across full-time workers. Table A-3 in the appendix present this ranking. Goos, Manning, Salomons (2014) use the 1993 wages across European countries (inclusive of Denmark) and arrive at a similar ranking as ours at the 2-digit ISCO occupations. 13 We follow the literature and focus on jobs outside of agriculture, however, the results are not affected by this. 14 Figure 3 is constructed using all employees of the manufacturing and service sectors respectively in 1999, who were born between 1945 and 1984. The figure shows the change in employment shares across high-, mid-level, and low wage occupations of these workers between 2000 and 2009.
10
The manufacturing sector employed 440,000 workers in Denmark in 1999, while in the same year 1,800,000 workers were employed in the service sector. The figure shows that manufacturing is much more strongly characterized by job polarization than services. In fact, employment growth in services is monotonically increasing with the wage, which is the pattern of the 1990s. As we will show below, trade does not only help to explain the U-shaped pattern in manufacturing but also the fact that employment growth in services is monotonically increasing across wage groups. This result confirms other findings that the growth of service jobs is to some extent the flip side of the hollowing out of mid-level wage jobs in manufacturing (Autor and Dorn 2013, Utar 2015). It is useful to keep in mind that our analysis tracks workers and their occupation, not their industry. When Figure 3 shows that manufacturing workers experience reductions in mid-level wage employment and increases in low-wage employment, in principle this could be because workers shift within manufacturing from mid-level to low-level jobs (within-industry shift), or because workers shift from mid-level occupations in manufacturing to low-wage occupations in other industries (between-industry shift). If employment shares were constant over time across industries, observing employment polarization in a particular industry such as textiles and clothing would imply that mid-level workers such as assemblers move in roughly equal parts into high wage and low wage jobs. While this is in principle possible, without between-industry shifts the potential for job polarization would necessarily seem to be limited: it is difficult to imagine that many machine operators and assemblers become professionals or managers while others become laborers, all in the same industry. Far from constant employment, however, the Danish textiles and clothing industry has been undergoing massive changes in terms of size: in the year 1999, it employed 13,000 workers, down from 25,000 in 1991, while by the end of 2010 industry employment was down to 4,500 workers.15 Part of this reduced demand for workers producing textiles in Denmark after 1999 is due to the removal of import quotas for China (Utar 2014), 15
In the year 1999, textiles and clothing accounted for 3 % of total manufacturing employment.
11
-.3
-.2
-.1
0
.1
.2
Changes in Occupational Employment Share, 2000-2009
Low-Wage
Mid-Wage
High-Wage
T&C Workers in 1999 Manufacturing Workers in 1999
Figure 4: Changes in employment shares of low-, mid-, and high wage occupations by workers’ initial industry, 2000-2009
although other factors, including a secular declining trend of the labor-intensive industries and technological factors, are surely present as well. Figure 4 shows the patterns of job polarization among textiles and clothing and manufacturing workers as a whole. We see that the job polarization pattern for textiles and clothing is stronger than for manufacturing as a whole. Thus, job polarization in textiles and clothing accounts over-proportionally for job polarization in manufacturing.16 In the following section we move from the aggregate to the micro level by providing some initial evidence on job polarization in Denmark based on movements between jobs by individual workers. 16
It is important to distinguish the impact of trade on job polarization at the firm- versus workerlevel. Figure 1 in the Supplementary Appendix show occupational employment share changes within the textile and clothing sector. In that figure, rather than following workers and their occupations, we follow employment within the T&C sector. Along the lines of Utar (2014), occupational employment share changes exhibit skill upgrading within the sector, more so among the firms that were exposed to imports coming from China.
12
3
Facts from individual worker job transitions
In this section we present several facts about worker transitions between occupations in Denmark during the sample period by looking at two specific two-digit occupations, drivers and mobile plant operators (ISCO 83) versus machine operators and assemblers (ISCO 82). Fact 1: There is substantial variation in worker outcomes at broadly similar wage levels but different occupations.
100 Drivers (ISCO 83) Machine Operators and Assemblers (ISCO 82) 90
80
Percent
70
60
50
40
30
20
2001
2002
2003
2004
2005
2006
2007
2008
2009
Year
Figure 5: RBTC: Probability of Staying for Machine Operators versus Drivers
Because both drivers and machine operators are among mid-level wage categories (see Table A-3), reduced employment in either of these occupations may help to explain the overall finding of job polarization. Nevertheless workers in these two occupations experienced differences in their job market experience over the sample period. Figure 5 shows the cumulative probability for a worker to stay in her initial occupation over 13
10 Drivers (ISCO 83) Machine Operators and Assemblers (ISCO 82)
9 8 7
Percent
6 5 4 3 2 1 0
2001
2002
2003
2004
2005
2006
2007
2008
2009
Year
Figure 6: The Rise of Service Worker Employment: Machine Operators versus Drivers
the period 2001 to 2009. First, note that there is a considerable amount of movement of these workers. By the year 2009, around two thirds of the workers do not work anymore in the occupation that they worked in the year 2001. This is to some extent the consequence of the relative ease of switching jobs (low hiring costs, low firing costs) in the Danish labor market. Second, over these eight years there are marked differences in the probability of staying in one’s initial occupation: in particular, drivers stay in their initial occupation with a 15 % higher probability than do machine operators and assemblers. In the textile and clothing sector, operators would attend fibre-preparing-, spinning-, and winding machines. They would also operate weaving, knitting, sewing, bleaching, and dying machines. Given the relatively narrow and programmable nature of their tasks it is plausible that the tasks of these operators and assemblers are prone to being automated
14
and done by robots. This is reflected in a relatively high ”routine task index” (RTI), which is 0.64 for this occupation.17 ,18 In contrast, drivers in the textile and clothing sector, who would drive cars, vans, trucks, as well as forklifts, perform generally less routine tasks. A driver of a truck takes his load to different destinations, diggers enter new terrain, or the personal driver of an executive makes new runs. The occupation of drivers has a relatively low routine task intensity (RTI of -0.99). Our finding that machine operators and assemblers are less likely to stay in their initial occupation than drivers points to the importance of routine-biased technical change (RBTC). Fact 2: Shrinking mid-level wage occupations are a source of growth of low-wage occupations, especially in services. Figure 6 shows the cumulative probabilities of machine operators and drivers to transition into personal and protective service occupations (ISCO 51). These service occupations include protective service workers as well as housekeepers, cooks, hairdressers, and travel guides. Service occupations require generally relatively few skills and wages are typically low. At the same time, service tasks are characterized by a relative low level of routinization (RTI of -0.23). Figure 6 shows that the cumulative probability that a machine operator transitions into these service jobs is increasing over time, and by the year 2009 7.5 % of all workers that were machine operators in 2001 have moved into person and protective service jobs. In contrast, drivers rarely move into these service worker jobs, and the cumulative probability is hardly increasing over time. This is consistent with recent evidence that the flip side of lower employment in high routine tasks due to RBTC is increasing service employment (Autor and Dorn 2013). Figure 6 confirms this result at the level of individual workers. Fact 3: Import competition leads to worker transitions consistent with job polarization. 17
See Autor, Levy, and Murnane (2003), Autor, and Dorn (2013). Autor (2013) discusses this tasks approach more broadly. 18 We thank Anna Salomons for sharing the RTI measure used in Goos, Manning, and Salomons (2014) with us.
15
100 Machine Operators with Trade Shock Machine Operators without Trade Shock Drivers with Trade Shock Drivers without Trade Shock
90
80
Percent
70
60
50
40
30
20
2001
2002
2003
2004
2005
2006
2007
2008
2009
Year
Figure 7: Import Competition and Job Polarization
In Figure 7 we show cumulative transitions of drivers (ISCO 83) and machine operators and assemblers (ISCO 82) out of their occupation depending on whether they were exposed to increased competition from Chinese producers, or not. We see that both drivers and machine operators are less likely to stay in their initial occupation in the presence of Chinese import competition.19 Over the entire period of 2001-2009, machine operators affected by Chinese import competition have a 29 % lower chance of staying in their initial industry compared to machine operators not affected by import competition. The corresponding figure for drivers is 22 %. In the light of our finding that workers highly exposed to RBTC (such as machine operators) have a 15 % higher chance to switch occupations than workers little exposed 19
The transitions of machine operators over time result in a smoother picture mainly because there are many more of them than drivers. In 2001 there were 2726 machine operators (ISCO 82) as opposed to only 101 drivers.
16
to RBTC (such as drivers), the difference in likelihood of staying associated with trade of 22-29 % appears to be sizable. In section 5 below we will return to this question with an approach that holds constant a range of worker and firm characteristics to estimate the causal effect of trade. Before we turn to this, the next section introduces our data and describes the policy change that is employed.
4
Workers facing import competition: the data
4.1
Employee-employer matched data
The main database used in this study is the Integrated Database for Labor Market Research, IDA, which is comprised of person, establishment, and job files.20 The person files contain annual information on all persons of age 15-70 residing in Denmark with a social security number. The establishment files contain annual information on all establishments with at least one employee in the last week of November of each year. The job files provide information on all jobs that are active in the last week of November in each year. IDA data-sets are complemented with the domestic production data-set (VARES) that covers all manufacturing firms with at least 10 employees, and the annual longitudinal data-set that matches firms with their employees (FIDA). The data-sets are from Statistics Denmark. For each worker we have information on their annual salary, hours worked, industry code of their primary employment, education level, demographic characteristics including age, gender and immigration status, and occupation of primary employment. Of particular interest to us here is the information on workers’ occupations. Occupational codes matter in Denmark because they influence a worker’s wage due to a collective bargaining system. 20
See Bunzel (2008) and Timmermans (2010) for more information.
17
Employers and labor unions pay close attention to occupational codes. This explains the administrative nature of the data. As a consequence, the quality of occupational data in Denmark is high compared to other countries.21 Information on workers’ occupation closely follows the international ISCO-88 classification. For most workers, occupational information is given at a detailed 4-digit level. There were close to 13,000 workers in Denmark’s textile and clothing industry in the year 1999. In our analysis we focus on workers who are in working age (17 to 67 years) throughout our sample period. So there are around 11,000 workers in our sample (see Table 1, Panel A for summary statistics). The textile and clothing industry is a typical manufacturing industry in which plant and machinery operator tasks (ISCO occupation code 8) play a large role; according to Figure A-1, they account for more than 40 % of all workers. Nevertheless the textile and clothing industry employs workers performing a diverse set of tasks. Technicians and associate professionals (ISCO 3), craft workers (ISCO 7), but also clerks (ISCO 4) account each for around 10 % of the textile and clothing workforce, whereas managers (ISCO 1) make up about 5 % (Figure A-1). Table A-1 presents worker characteristics for several occupation groups separately. We see that a majority of professionals (ISCO 2) in our sample is college educated (61 %) whereas, for example, college educated workers only constitute 6 % among craft workers (ISCO 7). Workers are on average 39 years old, indicating roughly the middle of the career span. The share of female workers is 57 %, and 6 % are immigrants (Table 1). The share of women is highest among clerks (78 %) and lowest among managers (21 %). Eleven percent of workers are college educated, and 35 % of workers have formal vocational training. In Denmark vocational education is provided by the technical high schools (after 9 years of mandatory schooling) and involves several years of formalized training including both schooling and apprenticeship. Roughly half of the workers have at most high school education. These figures show that our sample includes workers from a range 21
See, for example, Groes, Kircher, and Manovskii (2015) on this point.
18
of skill levels, important for our analysis of the effects of globalization on manufacturing workers in advanced countries. Compared to the economy as a whole these workers belong typically to the mid-level wage group. In Table 1 we report the breakdown of workers into a three-way wage classification of occupations. As reported in Table 1, 62 % of our workers held mid-level wage jobs in the year 1999. In the empirical analysis we also employ a number of additional variables to control for both worker and firm characteristics. They include worker experience, measured by the number of years each worker was in the labor force between 1980 and 1999, and the worker’s unemployment history between 1980 and 1999 (see Table 1). Also, note that 80 % of workers are labor union members, and close to 90 % are members in Denmark’s unemployment insurance program. The high degree of unionization is one indication that Denmark’s labor market institutions are quite different from those in the United States, for example, although it is worth noting that this does not make them necessarily more rigid. In particular, Denmark ranked 6th among 148 countries in terms of its hiring and firing practices, more de-regulated than the United States which came in 9th in the same ranking (Global Competitiveness Report 2014). Relatively unrestricted hiring and firing in Denmark is combined with active labor market policies and a generous social safety net to yield a system that has been termed “flexicurity”. While the comparatively limited wage flexibility in Denmark leads us to focus mostly on job polarization in terms of employment, we will show some results on job polarization in terms of earnings as well. The Danish production database is used to identify firms with domestic production in one or more of the goods that were subject to the MFA quotas for China. We identify firms that in 1999 produce 8-digit Combined Nomenclature (CN) level goods that are subject to the MFA quota removal for China, and using a firm identifier we map this information to worker-level information. In the year 1999, about half of the workers are exposed to increased import competition in the sense that they are employed at a firm that will subsequently be affected by quota removals as a consequence of China’s 19
accession to the WTO. Panel (b) of Figure A-1 compares the distribution of occupations in the exposed firms with that in the non-exposed firms as a sample balance check. We see that plant and machine operators is the largest occupation in both sets of firms, accounting for more than 40 % of the workforce in both exposed and non-exposed firms. The two sets of firms have also a similar share of managers (ISCO 1). There are some differences, for example exposed firms have a higher share of clerks (ISCO 4) than nonexposed firms, while non-exposed firms have higher share of craft workers (ISCO 7). Overall, however, we do not see major differences between the two sets of firms in panel (b) of Figure A-1. Given their importance in the textile and clothing industry, we examine the group of “plant and machine operators” occupations in more detail. Table A-2 reports summary statistics on workers initially employed at firms that were exposed to import competition, versus workers that were not exposed to import competition separately. For example, sewing machine operators in exposed firms in 1999 were 95.7 % women, while this fraction was 94 % at non-exposed firms. We also see that 4.3 % of weaving and knitting machine operators were college educated in non-exposed firms, while 4.9 % of these workers were college educated in exposed firms. The weaving and knitting machine operators at exposed firms had a labor market experience of about 16.5 years, compared to about 15.7 years in non-exposed firms. To the extent that longer labor market experience gives valuable skills, and firms do not want to let go relatively skilled workers, this would bias the analysis against finding an effect of trade on job polarization. The main finding from comparing the initial workforce in exposed and non-exposed firms though is that the differences are small and unlikely to explain our findings.
20
Table 1: Descriptive Statistics Panel A. Characteristics of Workers with Primary Employment in T&C in 1999 Mean SD Age 39.174 10.473 Female 0.571 0.495 Immigrant 0.061 0.240 College Educated 0.112 0.316 Vocational School Educated 0.350 0.477 At most High School 0.514 0.500 Years of Experience in the Labor Market 14.500 5.864 Unemployment History Index 127.241 182.741 Log Hourly Wage 4.968 0.348 High Wage Occupation 0.188 0.390 Mid Wage Occupation 0.622 0.485 Low Wage Occupation 0.112 0.315 Union Membership 0.800 0.400 Unemployment Insurance Membership 0.897 0.304
N 10753 10753 10753 10753 10753 10753 10753 10753 10753 10753 10753 10753 10753 10753
Panel B. Main Outcome Variables Cumulative Years of Employment in High- and Low Wage Jobs
-0.143
4.938
10753
net of Years of Employment in Mid Wage Jobs (2002-2009) Cumulative Hours Worked in High- and Low Wage Jobs net of
-0.062
6.573
10753
Hours Worked in Mid Wage Jobs (2002-2009) Cumulative Salary in High- and Low Wage Jobs net of Salary in
0.175
8.735
10753
Mid Wage Jobs (2002-2009) Cumulative Years of Employment in High Wage Jobs (2002-2009) Cumulative Years of Employment in Mid Wage Jobs (2002-2009) Cumulative Years of Employment in Low Wage Jobs (2002-2009)
1.379 2.548 1.025
2.505 2.847 1.991
10753 10753 10753
Notes: Variables Female, Immigrant, Union Membership, UI Membership, High Wage, Mid Wage and Low Wage Occupations, College Educated, Vocational School Educated and At most High School are worker-level indicator variables. The Unemployment History Index is the cumulative sum of the percentage of working time spent as unemployed within each year since 1980. Data Source: Statistics Denmark.
21
4.2
Import competition through quota removals
The Multi-fibre Arrangement was introduced in 1974 to govern world trade in the textiles and clothing. Under this agreement a large portion of textile and clothing exports from low-wage countries to more developed countries was subject to quantitative restrictions called quotas. The arrangement provided extensive protection for developed country textile and clothing industry against competition from low-wage country products. The liberalization of trade in the textile and clothing industry was not agreed until the conclusion of the Uruguay Round in 1995, when the Agreement on Textiles and Clothing (ATC) replaced the MFA, and provisions were made for phasing it out in four steps over a period of 10 years. Quotas were to be eliminated equivalent to 16 percent of 1990 imports at the beginning of 1995 (Phase I), 17 percent at the beginning of 1998 (Phase II), 18 percent at the beginning of 2002 (Phase III), and the remaining 49 percent at the beginning of 2005 (Phase IV). Quotas covered a wide range of both textile and clothing products ranging from bed linens to synthetic filament yarns to shirts but at the same time coverage within each broad product category varied, making it important to utilize MFA quotas at a detailed product-level. For example, “shawls and scarves of silk or silk waste” was part of a quota restriction for China while “shawls and scarves of wool and fine animal hair” was not. “brasseries of all types of textile materials” was under quota but not “corselettes of all types of textile materials”. As one of the smaller members of the EU, the extensive coverage of quotas was largely exogenous to the Danish industrial structure. Coverage of these quotas was determined throughout 1960s and 70s and the negotiations of the MFA were held at the EU level (Spinanger 1999). Since 1993 the quotas were also managed at the EU level. China did not become eligible for quota removal until it joined the WTO at the end of 2001. While there was a considerable amount of uncertainty as to whether China’s negotiations for WTO membership would succeed, we choose 1999 as the year to de22
termine whether a firm had a product in its portfolio that would be subject to a quota removal to limit any anticipation effects. In January 2002, its quotas on Phase I, II and III goods were removed immediately. By being a WTO member, China was also allowed to benefit from the scheduled last phase in January 2005.22 We utilize as our measure of import competition the abolishment of MFA quotas for China in conjunction with her accession to the WTO.23 Most of the quotas for China had more than 90 % filling rates. Figure 8 shows the evolution in textile and clothing import shares of China compared to other developing countries subject to MFA quotas. Clearly China is the country that stands out in this figure. Using transaction-level import data, Utar (2014) confirms that the MFA quotas were binding for China and both the 2002 and the 2005 quota lifting cause a substantial surge of MFA goods from China in Denmark with associated decline in unit prices of these goods.24 As a consequence, virtually all workers employed at firms subject to the quota removals faced increased import competition from China. Employing a twostage least squares strategy in which exposure to Chinese quota removal instruments for import competition generally leads to similar results (see Utar 2014). For this reason, here we employ exposure to Chinese quota removal as our treatment variable. 22
Due to a surge of Chinese imports in the first few months of 2005 at European Union (EU) ports in response to the final phase of the quota removal, the EU retained a few of the quota categories until 2008. Since the sample period extends over 2008, those few quotas are also included in the current analysis. 23 Here we focus on China’s accession to the WTO as a shock because while the ATC provided a schedule for gradual dismantling of MFA quotas already in 1995, removal of MFA quotas for China depended on whether and when it would join the WTO. Furthermore, there is significant overlap between firms producing quota products subject to 2002 or 2005 phases, making it difficult to separate the effects of different phases. Finally, due to a surge of Chinese imports in the first few months of 2005 at the EU ports in response to the final phase of the quota removal, the EU retained a few of the quota Phase IV categories until 2008. The selection of the retained categories were clearly endogenous to the EU-wide industrial structure as it was due to the pressure of the European industrialist. 24 Khandelwal, Schott, and Wei (2013) shows that the quota removal for China led to an extra efficiency gain in China due to prior mismanagement of quotas by the Chinese government and the decline in prices were a result of entry of more efficient Chinese producers into the export market.
23
China Argentina Brazil Hong Kong India Indonesia Korea Macao Malaysia Pakistan Peru Phillipines Singapore Sri Lanka Thailand
0.3
0.25
0.2
0.15
0.1
0.05
0 1999
2000
2001
2002
2003
2004
2005
2006
2007
2008
2009
Figure 8: Import shares of China and other developing countries subject to MFA quotas in Danish Textile and Clothing Imports 1999-2009
5
Empirical results
To examine the role of trade for job polarization, we consider the following regression:
JP emp = β0 + β1 T radei + ZiW + ZiF + i i
(1)
The dependent variable JP emp is a measure job polarization at the worker level (subscript i i) over the period from 2002 to 2009. On the right hand side we have the measure of trade (T radei ), as well as measures of worker (ZiW ) and firm (ZiF ) characteristics. The sample consists of all workers who are at the working age throughout 2002-2009 and had their primary employment in textiles and clothing in the year 1999. Equation (1) is a cross-sectional, worker-level regression relating job polarization during China’s post-
24
WTO membership period of 2002-2009 to worker-level characteristics in the pre-WTO year of 1999. In particular, job polarization is defined as JP emp i
=
T =2009 X
{Emphit + Emplit − Empm it }.
(2)
t=2002
In this expression, Empxit is a dummy variable that takes the value of one if worker i in year t (t = 2002, ..., 2009) has held a job in high-, mid-level, or low-wage occupations, denoted by x = h, m, l; our classification into these three brackets follows Table A-3.25 The dependent variable JPiemp is the number of years that worker i has held primary employment in a high-wage or low-wage occupation between 2002 and 2009, minus the number of years that worker i has worked primarily in mid-level wage occupations. By summing over high- and low-wage employment after netting out mid-level employment, JPiemp conveniently summarizes worker i’s polarization experience. Below we will also show results that treat the three wage groups separately. The variable T radei takes the value of one if in the year 1999 worker i is employed in a firm that domestically manufactures a product that with China’s entry into the WTO is subject to the abolishment of the MFA quotas for China, and zero otherwise. The vector ZiW includes the following controls for worker i in the year 1999: age, gender, immigration status, education level, and the logarithm of i’s average hourly wage.26 Further, ZiW includes initial occupation and education controls: indicators for worker i’s occupation in a high-, mid-level, or low-wage occupation in year 1999, and indicator variables for at least some college education, vocational education, and at most a (nontechnical) high school degree. We also include information on worker i’s labor market experience before 1999. First, this is the cumulative sum of the percentage of working time a worker spent as unemployed within each year since 1980, and the number of 25
We also use the European ranking of occupations as provided by Goos, Manning and Salomons (2004). It employs the two-digit International Standard Classification of Occupations 1988 (ISCO-88), and the ranking of occupations as high/mid-level/low is based on the 1993 mean European wage. Our results are robust alternatively using European-wide ranking. 26 We take the average wage for years 1999 and 2000 to smooth out temporary effects.
25
years of labor market experience before 1999. Finally, ZiW includes indicator variables for worker i’s union and unemployment insurance membership status in year 1999. The vector of firm controls ZiF holds constant differences across workers’ workplaces in the year 1999 that might induce a spurious correlation. It includes the size of the firm, measured by the number of full-time equivalent employees, firm quality, proxied by the logarithm of the average hourly wage paid in this firm, as well as a measure of the strength of the firm-worker bond, which is the percentage of workers that are not employed in the same firm from year 1998 to year 1999. In addition to the measure of job polarization based on years of employment, we will employ an alternative measure that captures changes in the hours worked of each worker. It is defined as
JPihrs
PT =2009 =
t=2002
{Hourshit + Hourslit − Hoursm it } Hoursit0
(3)
where Hoursxit is defined as the total number of hours worked by worker i in year t in high/mid-level/low-wage wage primary occupations. JPihrs captures the cumulative hours worked in high-wage jobs plus the hours worked in low- wage jobs, minus hours worked in mid-level wage jobs throughout 2002-2009 for worker i; the measure is normalized by Hoursit0 , which is worker i’s hours worked in his or her primary employment in initial year t0 .27 We also present results for an earnings polarization measure, defined as
JPiwage
PT =2009 =
t=2002
{Earningshit + Earningslit − Earningsm it } Earningsit0
(4)
where Earningsxit is total labor earnings of worker i from his/her high-, mid-level, or low-wage primary occupation, x = h, m, l. In the denominator of equation (4) are worker i’s initial earnings in his or her primary job. To reduce measurement errors, both JPihrs and the variable JPiwage , defined below, employ the average of the respective values for years 1999 and 2000. 27
26
5.1
Can trade explain the U-shaped job polarization pattern?
The estimate of β1 in equation (1) captures the impact of lifting the import quotas for China on Danish workers’ movements away from mid-level wage jobs towards highand low wage jobs. By focusing on the textiles and clothing sector we rule out that the results are affected by the secular decline of labor intensive industries in advanced countries, or the corresponding secular increase in non-manufacturing as these forces are in effect in both the control and the treatment group.28 In addition to industry we control for firm characteristics that could have an important influence on the results. Perhaps most importantly, the worker controls ensure that we compare workers within fairly narrow cells, making the workers that are exposed to import competition virtually identical to those that are not exposed to import competition from China. To see which set of variables matters we include them step by step. Table 2 presents the results from estimating equation (1) for the dependent variable JP emp . When only T radei is included on the right hand side, β1 is estimated at about 1.2, see column 1. The estimate indicates that workers exposed to import competition from China have more than one calendar year of additional employment in high- and low- wage jobs net of employment in mid-level wage jobs over the eight years from 2002 to 2009. The shift of employment from mid-level to low and high-wage jobs appears to be stronger for women than for men (column 2). Non-immigrants and relatively young workers exhibit the employment polarization more strongly than other workers. Including these demographic controls has only a small effect on the T radei estimate. If workers exposed to import competition from China held primarily mid-level jobs, our dependent variable would be relatively high for these workers simply because they disproportionately lose their jobs. To account for this we include indicator variables for worker i’s occupation in the year 1999 by wage categories (high-, mid-level, and low; 28
Figure A-2 shows that 52 % of all workers in the sample who were working in 2009 moved to the Service Sector by the end of 2009. Figure A-3 shows this figure separately for exposed and non-exposed workers. The figure is 60 % and 44 % respectively for exposed and non-exposed workers.
27
see column 3). We find that the indicators for high-wage and low-wage occupations in 1999 come in positive while the mid-level occupation indicator is negative; these results are plausible as they reflect the influence of the non-switchers on JP emp . Controlling for i initial occupations, we see that the T radei coefficient is now around 0.8.
We see further that college educated workers are more likely to be experiencing employment polarization while workers with vocational education less, although these coefficients are only weakly significant. Employment polarization is increasing in worker’s wage in 1999 and decreasing with labor market experience (columns 5 and 6). Union members and workers covered by unemployment insurance experience employment polarization less strongly than other workers, presumably because they switch out less of mid-level employment than other workers (column 7). Overall, beyond the occupation of a worker initially, controlling for additional worker characteristics has only minor effects on the T radei coefficient. We also add variables at the firm level on the right hand side of equation (1), see column 8. Firm heterogeneity might affect workers’ accumulation of human capital and may lead divergent labor market trajectories among workers. We find that polarization is less likely for workers employed in firms that pay relatively high wages. This may be both because workers prefer to not leave those firms and because such firms may have prepared themselves better than other firms for the onset of import competition from China. Furthermore, there is more polarization for workers at firms that typically have high separation rates, presumably because the firm-worker bond is less tight than in other firms. With all worker and firm controls included, the trade effect is estimated at around 0.8 with a robust t-statistic of about 10 (column 8). This means that on average exposed workers experience a gross churn of an additional 0.8 years of employment during 2002-2009, or 0.1 employment years for each year of the sample period. It is interesting to note that the coefficient estimate for female workers increases as other worker characteristics are controlled for. In column (7), the coefficient estimate 28
29
(2002-2009) (8) 0.816*** (0.089) 0.706*** (0.097) -0.606*** (0.181) -0.006 (0.005) 2.873*** (0.188) -1.659*** (0.162) 1.663*** (0.199) 0.579* (0.291) -0.589* (0.270) -0.443 (0.267) 1.283*** (0.154) -0.041*** (0.011) 0.000* (0.000) -0.311* (0.121) -0.621*** (0.159) -1.868*** (0.289) 0.000 (0.000) 0.015*** (0.002)
Notes: In all columns the number of observations is 10753. A constant is included but not reported. All control variables are belong to the initial sample period and part of ZiW or ZiF . Robust standard errors are reported in parentheses. ∗ , ∗∗ and ∗∗∗ indicate significance at the 10 %, 5% and 1% levels respectively. Data Source: Statistics Denmark.
Separation Rate
Firm Size
Dep. Var. : Cumulative Years of Employment in High- and Low Wage Jobs minus Years of Employment in Mid Wage Jobs (1) (2) (3) (4) (5) (6) (7) Trade 1.167*** 1.099*** 0.778*** 0.782*** 0.791*** 0.837*** 0.853*** (0.095) (0.096) (0.088) (0.087) (0.087) (0.088) (0.088) Female 0.346*** 0.668*** 0.629*** 0.789*** 0.758*** 0.820*** (0.097) (0.087) (0.087) (0.095) (0.095) (0.096) Immigrant -0.842*** -0.199 -0.326 -0.286 -0.624*** -0.505** (0.183) (0.169) (0.172) (0.173) (0.181) (0.182) Age -0.032*** -0.024*** -0.020*** -0.023*** -0.002 -0.003 (0.004) (0.004) (0.004) (0.004) (0.005) (0.005) High-Wage Occ 3.179*** 2.856*** 2.650*** 2.719*** 2.856*** (0.172) (0.178) (0.188) (0.189) (0.188) Mid-Wage Occ -2.102*** -2.088*** -2.134*** -2.025*** -1.757*** (0.154) (0.153) (0.154) (0.156) (0.161) Low-Wage Occ 1.240*** 1.290*** 1.288*** 1.348*** 1.560*** (0.194) (0.194) (0.193) (0.195) (0.197) College 0.621* 0.566 0.536 0.580* (0.294) (0.293) (0.292) (0.292) Vocational -0.651* -0.645* -0.569* -0.525 (0.272) (0.271) (0.271) (0.271) at most High School -0.466 -0.421 -0.362 -0.372 (0.270) (0.269) (0.268) (0.268) Log Hourly Wage 0.632*** 0.823*** 0.908*** (0.140) (0.144) (0.144) Experience -0.062*** -0.047*** (0.011) (0.011) Unemployment History 0.000* 0.001** (0.000) (0.000) Union Membership -0.354** (0.121) UI Membership -0.620*** (0.159) Firm Wage
Table 2: Job Polarization and Trade - Years of Employment
for the Female indicator is 0.8 and significant at the 1 percent level, that is women have an additional of 0.8 years of employment away from mid-level jobs toward the tails of the distribution. The major change occurs when we distinguish movers from stayers by including the initial occupation controls (compare columns 2 and 3). We will return to this issue in the discussion below. Once firm-level controls are included, the coefficient decreases a bit to 0.7, this indicates that women were disproportionately employed in firms that are more prone to job polarization, with lower wages and higher separation rates. Table 3 presents analogous results for our cumulative hours worked measure, JPihrs . While the findings are generally similar to those with years of employment, the tendency towards job polarization is somewhat stronger for hours than for years of employment (the T radei coefficient β1 is estimated at 0.87 with hours (column 8), compared to 0.82 with years of employment). This indicates that reducing mid-level worker hours, even if they continue to be employed, is another channel through which import competition generates job polarization. But the magnitude differences implies that this channel is relatively unimportant in Denmark.29
We find also a significant effect from import competition on cumulative earnings polarization, see Table 4. The estimate of 0.96 (column 8) implies that the average exposed worker sees a shift in earnings away from mid-level and towards high- and low-wage earnings that is roughly equal to the worker’s annual earnings before the fall of the import quotas for China. The coefficient estimate implies an annualized shift of roughly 12 % of one year’s earnings over the period of 2002 to 2009. This is quite similar to the average annual effect measured in hours worked, 11 % of initial annul hours worked as well as the effect measured in years worked, 10 % of a year. Import competition has a sizable impact on (cumulative) earnings polarization because workers move across occupations. 29
Utar (2015) notes that the main channel through which the import shock is felt at the initial employer is a reduction in tenure at the initial employer, rather than reduction in hours worked or the hourly wage.
30
31
in High- and (2) 1.090*** (0.128) yes
Low Wage Jobs minus Hours Worked in Mid Wage Jobs (3) (4) (5) (6) 0.776*** 0.781*** 0.789*** 0.861*** (0.121) (0.120) (0.120) (0.122) yes yes yes yes yes yes yes yes yes yes yes yes yes yes
(2002-2009) (7) 0.882*** (0.122) yes yes yes yes yes yes
(8) 0.867*** (0.124) yes yes yes yes yes yes yes
Notes: In all columns the number of observations is 10753. A constant is included but not reported. Demographic controls are age, gender dummy, and immigration status dummy. Education controls are dummy variables indicating whether an individual has at most high school degree, vocational training (after high school) or college and above degree in 1999. Labor Market History variables are unemployment history and experience variables. Unemployment history is the number of years between 1980-1999 that the individual spent as an unemployed person. Experience is the number of years person was in the labor market between 1980-1999. Initial wage is the logarithm of the average hourly wage of an individual (from his/her primary occupation in T&C) in 1999 and 2000. Initial Occupations are indicator variables whether an individual was employed in 1999 having a high-wage, mid-wage or low-wage occupation (outside category is unspecified occupations). Union and UI Controls are Union membership and UI membership variables. Union membership is a dummy variable indicating if the individual is a member of a labor union. UI membership is a dummy variable indicating if the individual is a member of the (voluntary) Unemployment Insurance Fund. Initial workplace controls are the logarithm of the average hourly wage in the workplace in 1999, the number of full-time equivalent number of employees in 1999, and the separation rate in 1999 (percentage of employees that left the workplace since the previous year (1998)). Robust standard errors are reported in parentheses. ∗ , ∗∗ and ∗∗∗ indicate significance at the 10 %, 5% and 1% levels respectively. Data Source: Statistics Denmark.
Dep. Var. : Cumulative Hours Worked (1) Trade 1.180*** (0.126) Demographic Controls Initial Occupations Education Controls Initial Wage Labor Market History Union and UI Controls Initial Firm Controls
Table 3: Job Polarization and Trade- Hours Worked
32
Jobs minus Primary Earnings in Mid Wage Jobs (2002-2009) (4) (5) (6) (7) (8) 0.896*** 0.877*** 0.948*** 0.981*** 0.955*** (0.162) (0.164) (0.169) (0.169) (0.171) yes yes yes yes yes yes yes yes yes yes yes yes yes yes yes yes yes yes yes yes yes yes yes yes yes
Notes: In all columns the number of observations is 10753. A constant is included but not reported. Demographic controls are age, gender dummy, and immigration status dummy. Education controls are dummy variables indicating whether an individual has at most high school degree, vocational training (after high school) or college and above degree in 1999. Labor Market History variables are unemployment history and experience variables. Unemployment history is the number of years between 1980-1999 that the individual spent as an unemployed person. Experience is the number of years person was in the labor market between 1980-1999. Initial wage is the logarithm of the average hourly wage of an individual (from his/her primary occupation in T&C) in 1999 and 2000. Initial Occupations are indicator variables whether an individual was employed in 1999 having a high-wage, mid-wage or low-wage occupation (outside category is unspecified occupations). Union and UI Controls are Union membership and UI membership variables. Union membership is a dummy variable indicating if the individual is a member of a labor union. UI membership is a dummy variable indicating if the individual is a member of the (voluntary) Unemployment Insurance Fund. Initial workplace controls are the logarithm of the average hourly wage in the workplace in 1999, the number of full-time equivalent number of employees in 1999, and the separation rate in 1999 (percentage of employees that left the workplace since the previous year (1998)). Robust standard errors are reported in parentheses. ∗ , ∗∗ and ∗∗∗ indicate significance at the 10 %, 5% and 1% levels respectively. Data Source: Statistics Denmark.
Dep. Var. : Cumulative Primary Earnings in High- and Low Wage (1) (2) (3) Trade 1.306*** 1.200*** 0.891*** (0.168) (0.171) (0.163) Demographic Controls yes yes Initial Occupations yes Education Controls Initial Wage Labor Market History Union and UI Controls Initial Firm Controls
Table 4: Job Polarization and Trade - Earnings
5.2
Trade and workers’ shift across occupations with different wage levels
While the previous section has shown that trade can explain the U-shaped pattern of job polarization, we have so far employed a summary measure of polarization, additions to the tails minus reductions in the middle of the wage distribution. In order to see what the relative contribution of tails versus middle of the distribution is, and whether there is indeed evidence that the hollowing out of the middle is the flip side of gains in the tails, we now turn to analyzing the high-, mid-level, and low-wage occupations separately. The regression specification is analogous to equation (1), except that the dependent variables are summing the years of employment for each occupation category x = h, m, l separately. For example, JP h,emp is the number of years worker i was employed in highi wage occupations during the years 2002 to 2009, and analogously JP m,emp and JP l,emp i i for mid-level and low-wage employment.
Table 5 presents these results. We see that workers exposed to import competition from China spend about 0.2 years more on average in high-wage jobs, compared to otherwise similar workers that are not exposed to import competition from China. For mid-level employment, the coefficient on T radei is about −0.4, while for low-level employment the coefficient is again about 0.2 (columns 2 and 3, respectively). Trade has hollowed out the middle of the wage distribution, and it has added jobs both at the top and the bottom of the distribution. Furthermore, the sum of the T radei coefficients across the occupation categories is approximately equal to zero, and it is likely that the fatter tails and the hollowed out middle are the flip sides of the same phenomenon. Additional analysis shows that hours worked losses in mid-level wage jobs are only to 2/3 compensated by gains in high-wage and low-wage hours, and high-wage hour gains are more important than low-wage gains. Results for earnings polarization are similar (results are available
33
34
0.229*** (0.040) 0.199*** (0.045) no
-0.394*** (0.052) -0.335*** (0.057) no
JP m,emp i (2) 0.193*** (0.038) 0.171*** (0.041) no
JP l,emp i (3) 0.216*** (0.040) 0.032 (0.046) yes
JP h,emp i (4)
-0.382*** (0.052) -0.371*** (0.060) yes
JP m,emp i (5)
0.178*** (0.038) 0.292*** (0.043) yes
JP l,emp i (6)
Notes: In all columns the number of observations is 10753. Although only coefficient estimates of ‘Female’ control is reported, all regressions include a constant and the full-set of controls, ZiW and ZiF . In columns (4)-(6) initial occupations are controlled for using 2-digit ISCO occupation dummies. Robust standard errors are reported in parentheses. ∗ , ∗∗ and ∗∗∗ indicate significance at the 10 %, 5% and 1% levels respectively. Data Source: Statistics Denmark.
2-Digit ISCO Dummies
Female
Trade
JP h,emp i (1)
Table 5: Job Polarization and Import Competition - Years of Employment in High, Mid, and Low Wage Jobs
upon request).30 We have also confirmed the robustness of the results by employing a logarithmic transformation of the dependent variable, as well as the revenue share of MFA goods instead of the T radei indicator variable. See Tables A-4 and A-5 in the appendix.31
5.2.1
Worker’s susceptibility to job polarization
Occupation While the set of regressors includes indicator variables for each worker’s initial occupation in the broad wage distribution (high, mid-level, or low-wage occupation), there are differences across occupations within broader occupation groups in how susceptible workers are to job polarization; we have seen evidence of that in the case of machine operators versus drivers in section 2 above. If this variation would be correlated with exposure to import competition our results might be spurious, even though Figure A-1 shows that the distribution of broad occupational categories for workers at trade exposed and non-exposed firms was quite similar. We address these concerns by adding indicator variables for worker i’s employment in each of the 22 two-digit occupational codes. The results for high-, mid-level, and low-wage employment job polarization regressions turn out to be not strongly affected by the inclusion of the two-digit occupational dummies. The trade coefficients for high-, mid-level, and low-wage employment remain roughly at 0.2, −0.4, and 0.2, respectively (columns 4, 5, and 6 of Table 5). From this it does not appear to be the case that the trade effect on job polarization is driven 30
Our findings here are in line with Utar’s (2015) analysis of worker’s adjustment costs. She shows that workers who move down the ladder of occupations to ’low-end’ service sector jobs face significant adjustment costs in the form of difficulties in keeping these jobs. These workers suffer from frequent unemployment disruptions while their primary attachment to the labor market continued to be these ’low-end’ jobs. 31 In Table A-4 we provide the results from transforming the dependent variables into logarithmic scale (after adding an arbitrary amount to make the minimum value one) and the results are robust. We also employ the manufacturing revenue share of goods that were subject to the removal of quotas for China in 1999 in worker i’s initial employer as an alternative trade exposure variable. See Utar (2014, 2015) on these modifications. The revenue share trade variable, T^ radei , takes into account the degree in which the initial firms were exposed to the competition. We find that our results are robust (see Table A-5).
35
by differences in the susceptibility of occupations for polarization. In addition, we have examined a number of major occupations in more detail by creating interaction variables, such as T radei × M anagersi , where M anagersi is equal to one if worker i worked as a manager in year 1999, and zero otherwise. Other occupations for which we construct interaction variables are professionals, associate professionals and technicians, clerks, and machine operators. These interaction variables are introduced together with the T radei variable. We find that plant and machine operators (ISCO 8) constitute the group that is most strongly behind the result that trade causes job polarization (Table A-6, Panel F). We also present the analysis separately only among machine operators (ISCO 82) (see Table A-7) and find that trade leads to a loss of eight months of mid-wage employment, and increases of low-wage and high-wage employment of about four months and one month respectively over a period of eight years. Outside of machine operator occupations, trade does in fact not cause a significant hollowing out of employment in mid-level wage jobs. Managers, associate professionals and technicians, and clerks hardly contribute to our finding that trade causes job polarization (Panels A, C, and D). Interestingly, professionals significantly gain mid-level employment through trade, which is the opposite of the experience of machine operators (Panel B). If job switching triggered by trade lets professionals to disproportionately stay within the same sector as their tasks are relatively demanded compared to other occupations (see Utar 2014, 2015 as well as Figure 1 in the Supplementary Appendix), this combined with the fact that manufacturing sector is relatively rich in mid-wage jobs compared to service sectors may explain the difference. This provides initial evidence for significant differences in the impact of trade across different occupations. We will return to this in more detail when discussing the relationship between trade and task characteristics in contributing to job polarization in section 5.4 below.
36
Education We are also interested in how workers with different levels of education fare in the face of import competition from China. We employ another set of interaction variables of the form T radei × Educi , where Educi takes on a value of one if the highest education level of worker i in 1999 is college, vocational school, and at most high school, respectively, and zero otherwise. Including the trade-education interaction together with the T radei variable, we find that college-educated workers do not contribute to job polarization caused by trade (Table A-8, Panel A), in fact college-educated workers at firms exposed to import competition tend to gain, not lose, in terms of mid-level employment. In contrast, workers with at most high school education bear the brunt of the job polarization caused by trade (Panel C). This is in line with the general pattern of job polarization observed in Denmark between 2000-2009 (see Figure 4 in the Supplementary Appendix).
5.2.2
Job polarization through shifts within versus between industries
In this section we discuss the impact of trade for high-, mid-level, and low-wage employment across sectors. The results are shown in Table 6. In the upper left corner, recall that the coefficient on T radei of 0.229 means that workers employed in 1999 in textiles and clothing firms exposed to import competition from China have had between 2002 and 2009 0.229 years of employment more in high-wage occupations than workers that were not employed by exposed textiles and clothing firms. The question addressed in this section is whether the job polarization impact of trade is associated with worker movements between broad sectors. As columns 2 and 3 show, the gain in high-wage employment that trade has caused is entirely due to workers moving into sectors outside of manufacturing (coefficient of 0.283). There is no significant high-wage employment gain through trade within manufacturing (coefficient of -0.054).32 On the other hand, the employment loss in mid-level jobs is 32
By construction, the coefficient in column 1 is the sum of the coefficients in columns 2 and 3, and analogously for the other columns of Table 6.
37
entirely due to reductions in mid-level manufacturing jobs (columns 5 and 6). Indeed trade causes increase in mid-wage jobs in the service sector (as it pushes workers toward the service sector), yet the net effect is negative. Columns 8 and 9 show that the increase in low-wage employment is split between manufacturing jobs and non-manufacturing. Overall, the pattern can be summarized by noting that the mid-wage employment losses are due to decline in manufacturing jobs while employment gains at the tails require to leave manufacturing, especially if a worker succeeds in moving to high-wage employment. This finding shows that frictions to the movement of workers across sectors implies that international trade can increase welfare for some workers (those with sufficiently low cost of moving) while at the same time it decreases welfare of other workers (those with high cost of moving), irrespective of any other effects of trade on welfare. Because non-manufacturing includes a broad set of diverse activities, in the following we isolate the services sector, see Panel B of Table 6.33 The results indicate that the employment changes caused by trade into non-manufacturing are almost entirely accounted for by shifts into the services sector.34 Specifically, columns 1-3 of Panel B show that 100 % of all high-wage employment gain is due to service sector jobs and columns 8 through 9 shows that 60 % of all low-wage employment gain is due to ’low-paid’ service jobs. Earlier work has emphasized the service sector in the context of job polarization (Autor and Dorn 2013). We extend this evidence by showing on the basis of micro evidence that this pattern is particularly true for job polarization caused by trade.
33 34
The other non-manufacturing sectors are agriculture, fishing, energy and construction. See also Utar (2015) on this.
38
39
0.229*** (0.040) 0.199*** (0.045)
JP h,emp i
0.229*** (0.040) 0.199*** (0.045)
JP h,emp i
(1)
JP h,emp i Within Manuf. -0.054 (0.034) 0.049 (0.035)
JP h,emp i Within Manuf. -0.054 (0.034) 0.049 (0.035)
(2)
JP h,emp i Within Service 0.276*** (0.030) 0.158*** (0.034)
JP h,emp i Outside Manuf. 0.283*** (0.030) 0.150*** (0.034)
(3)
-0.394*** (0.052) -0.335*** (0.057)
JP m,emp i
-0.394*** (0.052) -0.335*** (0.057)
JP m,emp i
(4)
JP m,emp i Within Manuf. -0.551*** (0.048) -0.339*** (0.053)
JP m,emp i Within Manuf. -0.551*** (0.048) -0.339*** (0.053)
(5)
JP m,emp i Within Service 0.156*** (0.027) 0.128*** (0.031)
JP m,emp i Outside Manuf. 0.157*** (0.029) 0.004 (0.034)
(6)
0.193*** (0.038) 0.171*** (0.041)
JP l,emp i
0.193*** (0.038) 0.171*** (0.041)
JP l,emp i
(7)
JP l,emp i Within Manuf. 0.074** (0.023) -0.120*** (0.025)
JP l,emp i Within Manuf. 0.074** (0.023) -0.120*** (0.025)
(8)
JP l,emp i Within Service 0.116*** (0.031) 0.332*** (0.032)
JP l,emp i Outside Manuf. 0.119*** (0.032) 0.291*** (0.033)
(9)
Notes: In all columns the number of observations is 10753. A constant is included but not reported. Although only coefficient estimates of ‘Female’ control is reported, all regressions include a constant and the full-set of controls, ZiW and ZiF . Robust standard errors are reported in parentheses. ∗ , ∗∗ and ∗∗∗ indicate significance at the 10 %, 5% and 1% levels respectively. Data Source: Statistics Denmark.
Female
Trade
Panel B. Dep. Var
Female
Trade
Panel A. Dep. Var
Table 6: Job Polarization and Job Shifts Within versus Outside Manufacturing
5.3
Technology and offshoring explanations for job polarization
In existing analyses of the causes of job polarization technology explanations play a central role. We therefore expect that routine-biased technical change (RBTC) helps to explain job polarization in our sample. The question is only to what extent, and whether trade mimics RBTC, in which case after controlling for RBTC the T radei variable will lose its significance. We are also interested in the extent to which the import competition effect is correlated with offshoring. China is a major offshoring destination, especially after it entered the WTO, and an increase of textiles and clothing goods from China could be related to offshoring decisions of Danish firms. Alternatively, if trade turns out to be unrelated to offshoring, it might be more plausible to think of autonomous Chinese production as the result of technology diffusion to Chinese-owned firms. We employ measures of RBTC and offshoring that are well established in the literature. One, the potential for occupational changes due to RBTC is captured by the routine task intensity (RTI; the variable is denoted as RT Ii ) introduced in section 2. Two, the likelihood of polarization due to offshoring, denoted Of f shoringi , is captured by a measure of Goos, Manning, and Salomons (2014).35 Both RBTC and offshoring variables are defined at the two-digit level of occupations.36 Table 7 presents these results. All regressions include the full set of worker and firm controls of Table 2. We see that workers who were exposed to the import shock have a significantly higher likelihood of moving to jobs that are located at either end of the wage distribution, while differences in offshorability help to explain job polarization as well (column 1). The table gives standardized beta coefficients in square brackets, which indicate that trade has a larger effect on job polarization than offshoring. Given the central role of computers for technical change in manufacturing recently, we add next a measure of the intensity of the 35
We have also considered Blinder and Krueger’s (2013) offshoring measure, finding broadly similar results. 36 Note that RBTC and offshoring measures are taken from Goos, Manning, and Salomons (2014) and they are at the 2-digit ISCO level.
40
worker’s interaction with computers (ICTi , column 2). Note that this does not simply measure whether the worker is prone to be replaced by a computer. Rather, the measure picks up that there are certain aspects of the job that can be codified, explaining the presence of the computer, but the worker’s task might involve to monitor and evaluate output from the computer, which is likely a non-routine task. We find no significant effect for this ICT variable. However, when we add the routine task index (RTI), not only is there more job polarization for routine tasks, as expected, but interacting with computers now enters with a negative sign (column 3). This suggests that the routine task index helps to separate routine from non-routine tasks in our ICT variable. The positive and significant coefficient on RTI in column 3 confirms earlier results that RBTC is a factor explaining job polarization in Denmark. We estimate a coefficient for T radei that is hardly affected by the inclusion of the offshoring and task variables. In terms of beta coefficients the impact of trade appears to be substantial, larger than that of technology factors captured by the RTI variable. Next, we investigate the relationship between import competition from China and offshoring. While offshoring affects primarily intermediate inputs trade and the China quota removal leads to an inflow of final goods into Denmark (Utar 2014), it is nevertheless possible that the two are related. In particular, it could be that certain tasks that workers used to perform in Denmark are offshored to China, where they are combined with other tasks before the final textiles and clothing good is exported from China to Denmark.
We employ an interaction variable between import competition from China and offshoring, T radei × Of f shoringi , to see whether there is evidence for a complementary relationship between offshoring and imports from China. The inclusion of the interaction variable turns the linear offshoring effect to zero, while the size of the linear T radei effect shrinks to around 0.3 (column 4). Offshoring and imports from China appear to 41
42 no
0.839*** (0.098) [0.083] 0.145** (0.044) [0.046]
no
0.842*** (0.098) [0.084] 0.138** (0.047) [0.043] -0.029 (0.069) [-0.006]
JPiemp (2)
no
0.833*** (0.098) [0.083] 0.135** (0.047) [0.043] -0.226** (0.088) [-0.044] 0.348** (0.106) [0.060]
JPiemp (3)
no
0.341** (0.131) [0.034] 0.002 (0.053) [0.001] -0.213* (0.088) [-0.041] 0.369*** (0.107) [0.064] 0.335*** (0.059) [ 0.087]
JPiemp (4)
ZiW
no ZiF ,
-0.654*** (0.100) [ -0.090]
0.712*** (0.100) [0.071] 0.116* (0.047) [0.037] 0.073 (0.098) [0.014] 0.381*** (0.107) [0.066]
JPiemp (5)
-0.026 (0.115) [ -0.004] no
0.848*** (0.114) [0.084] 0.135** (0.047) [0.043] -0.224* (0.088) [-0.044] 0.359** (0.116) [0.062]
JPiemp (6)
yes
0.769*** (0.104) [0.076]
JPiemp (7)
Notes: In all columns the number of observations is 8702. A constant and the full set of controls, and are included in all regressions but not reported. “Offshoring” is the offshorability index of the corresponding two digit initial ISCO-88 occupation code of a worker. It is constructed by Goos, Manning and Salomons (2014). “ICT” is the index variable indicating the degree in which the initial (4-digit) occupation of a worker interacts with computers. It is an O*NET (version 14) variable and varies across 4-digit occupations. “RTI” is the routine intensity index of the corresponding two digit initial ISCO-88 occupation code of a worker. The RTI index follows Autor, Levy and Murnane (2003) and Autor and Dorn (2012). Offshoring index is not available for certain ISCO 2-digit categories (see the Supplementary Appendix) and ICT index is not available across certain 4-digit ISCOs as they are provided by O*NET. Finally 7 % of workers have undetermined initial occupations. Hence running the estimations in a common sample decreases the number of observations to 8702. Control variables are described in the notes of Table 3. Standardized coefficients are reported in square brackets. Robust standard errors are reported in parentheses. ∗ , ∗∗ and ∗∗∗ indicate significance at the 10 %, 5 % and 1 % levels respectively. Data Source: Statistics Denmark.
4-Digit ISCO Dummies
Trade*RTI
Trade*ICT
Trade*Offshoring
Routine Task Intensity (RTI)
Interacting with Computers (ICT)
Offshoring
Trade
JPiemp (1)
Table 7: Job Polarization, Offshoring, Technology, and Trade
be complements, most likely because task offshoring leads to new final goods that are being shipped from China to compete for customers in Denmark. We also consider an analogous interaction variable between T radei and ICTi ,. Trade hardly leads to job polarization through workers whose tasks involve interacting with computers, see column 5. When instead we include an interaction between T radei and the broader RT Ii variable, it does not enter significantly (column 6). It appears that the ICTi variable captures quite well the task characteristics that shield a worker from job polarization.37 Finally, column 7 introduces four-digit occupational indicator variables. By construction, the occupational dummies fully account for the variation in RTI–our RTI variable varies at the two-digit ISCO level–, and the trade effect is with 0.77 similar in size to before.
5.4
Unpacking the trade effect
It is important to go further because as noted above, it has been challenging at times to distinguish trade from technology causes of job polarization. In this section we look at the intersection of exposure to import competition and specific tasks from the O*NET data base at the level of individual workers.38 Our approach will be to interact the 0/1 variable T radei with a measure of the importance of a particular task in that worker i’s four-digit occupation class. The regression of cumulative worker employment in 2002 to 2009 then includes three terms, in addition to all worker and firm characteristics from before. These three variables are the linear trade and task measures, denoted by T radei and ON ETi , together with their interaction T radei × ON ETi . In Table 8 we show results for six specific O*NET measures; additional results can be found in Tables 2 to 37
As we will see below, workers interacting with computers are less likely to lose mid-level wage jobs, and also less likely shift to low-wage jobs than the average worker (Table 8). 38 We use O*NET June 2009 version 14. See the data appendix section in the Supplementary Appendix for further details.
43
8 of the Supplementary Appendix.39 The six O*NET tasks are representative of routine manual tasks (Panel A), non-routine manual tasks (Panel B), routine cognitive tasks (Panel C), abstract tasks (Panel D), information and communication technology-intensive tasks (Panel E), and tasks that are intensive in face-to-face communication (Panel F). Tasks along these lines have been emphasized in influential studies of job polarization, and based on this work one would expect that routine tasks contribute to job polarization (Panes A and C) while nonroutine tasks do not (Panels B and D) (Autor, Levy, Murnane 2003, Autor and Dorn 2013). Tasks that involve intensively interacting with computers would also tend to be non-routine (the computer itself performs the routine part of the tasks), while faceto-face communication intensive tasks tend to prevent offshoring and thus reduce job polarization (Blinder and Krueger 2013, Firpo, Fortin, and Lemieux 2011). We present results both for the measure that sums the absolute value of employment changes in high-, mid-level, and low-wage jobs (JPiemp ), and for each of these three wage groups separately. Starting with the former, we see that workers have been affected quite differently by trade depending on the specific tasks they perform. On average trade has caused an employment churn of about 0.8 years in our JPiemp measure. The first column of Table 8 shows that the marginal effect of T radei ranges from essentially zero for routine cognitive tasks (0.524+(-0.587)= -0.063, see Panel C) to a whopping 1.4 years in the case of nonroutine manual tasks such as those requiring gross body coordination (Panel B). At the same time, while there is variation in the size of trade’s impact on job polarization, trade has contributed to job polarization through workers performing most kinds of tasks. We also see that trade’s effect is particularly high for manual tasks. In contrast, as expected there are tasks (and associated technologies) that contribute to 39
In order to keep the analysis transparent we utilize individual O*NET measures and check the robustness of the results by employing similar O*NET measures classified under the same broad group, rather than creating new composite indices; on this point, see Autor (2013).
44
job polarization (a positive effect on JPiemp ) while other tasks are associated with lower levels of job polarization (a negative effect on JPiemp ). The marginal effect of ON ETi ranges from 0.31 (-0.30+0.61=0.31, Panel B for non-routine manual tasks) to -0.42 (for ICT intensive tasks Panel E). While routine manual tasks are prone to polarization, we do not find evidence that this is the case for routine cognitive tasks.40 In line with expectations is the result that tasks that involve interacting with computers and faceto-face tend to be less associated with job polarization (Panels E and F). Looking at the three individual wage groups, we see that across different tasks trade has caused between almost 0.8 and virtually 0 years of employment less in the midlevel wage occupations (Panels B and E, column 3). While on average trade leads to a higher employment of 0.2 years for high-wage occupations, workers performing routine cognitive tasks do not experience that (column 2, Panel C). In contrast, trade has a particularly strong effect of shifting workers performing manual tasks into low-wage occupations (column 4). At the same time, workers that perform ICT-intensive tasks or require face-to-face communication are less likely shifted through trade into low-wage jobs (column 4, Panels E and F). Overall, there is evidence that technology factors and trade overlap to some extent in causing job polarization. Manual tasks, in particular, whether routine or non-routine, are prone to job polarization both through technology factors and trade. The fact that T radei and ON ETi variables are often both significant in Table 8 provides evidence that technology and trade are sufficiently distinct that we can separately identify their effect on labor demand. It is also useful to contrast our trade and technology measures with each other. The O*NET measures highlight important aspects of the worker’s tasks. At the same time, while the variable Repetitive Motions suggests that these tasks could be replaced by a robot (e.g. an automated machine-driven process controlled by a computer), it is not direct evidence of robotization, and it is hard to delineate the 40
This is confirmed by results for additional O*NET categories, see Tables 2 and 4 in the Supplementary Appendix.
45
factors that the variable Repetitive Motions picks up. Our comparison of workers who suddenly are or are not exposed to new import competition from China makes for a comparatively easy interpretation of the coefficient on T radei . Strong identification of the impact of trade is also suggested by the fact that the T radei × ON ETi interaction effect is estimated quite precisely in all six panels of Table 8 (column 1).41 We now examine gender differences in the way that trade has caused job polarization in Denmark.
41
Also note that in many cases it is the trade-task interaction that yields what one would generally expect as the task measure’s marginal effect on job polarization based on earlier research.
46
Table 8: Tasks, Trade, and Polarization Dep. Var. Panel A. Trade Repetitive Motions Trade*Repetitive Motions Panel B. Trade GBC Trade*GBC Panel C. Trade Evaluating Trade*Evaluating Panel D. Trade Thinking Trade*Thinking Panel E. Trade ICT Trade*ICT Panel F. Trade Interpersonal Trade*Interpersonal
JPiemp JPih,emp JPim,emp JPil,emp (1) (2) (3) (4) Routine Manual Tasks Tasks with Repetitive Motions 0.570*** 0.270*** -0.229*** 0.071 (0.117) (0.066) (0.062) (0.043) -0.224* -0.312*** 0.050 0.137*** (0.089) (0.041) (0.053) (0.034) 0.397*** -0.022 -0.274*** 0.145*** (0.100) (0.052) (0.055) (0.038) Non-Routine Manual Tasks Tasks with Gross Body Coordination (GBC) 0.792*** 0.196*** -0.399*** 0.197*** (0.095) (0.043) (0.055) (0.040) -0.304*** -0.494*** 0.050 0.240*** (0.091) (0.047) (0.052) (0.030) 0.610*** 0.073 -0.374*** 0.163*** (0.114) (0.058) (0.064) (0.040) Routine Cognitive Tasks Evaluating Information to Determine Compliance with Standards 0.524*** 0.123* -0.281*** 0.120** (0.107) (0.054) (0.060) (0.043) 0.289*** 0.019 -0.130** 0.139*** (0.081) (0.038) (0.048) (0.031) -0.587*** -0.152** 0.263*** -0.171*** (0.113) (0.056) (0.064) (0.047) Abstract Tasks Tasks that involves thinking creatively (Thinking) 0.633*** 0.253*** -0.314*** 0.066 (0.113) (0.063) (0.061) (0.045) 0.200* 0.123** -0.070 0.008 (0.081) (0.038) (0.048) (0.034) -0.280** 0.038 0.137* -0.181*** (0.098) (0.054) (0.053) (0.040) ICT Tasks that involves interacting with computers (ICT) 0.669*** 0.179*** -0.330*** 0.160*** (0.096) (0.048) (0.054) (0.038) 0.200* 0.359*** 0.009 -0.150*** (0.080) (0.038) (0.047) (0.029) -0.641*** -0.014 0.387*** -0.239*** (0.097) (0.048) (0.055) (0.037) Face to Face Communication Intensive Tasks Tasks that involve establishing interpersonal relationships 0.526*** 0.230*** -0.221*** 0.076 (0.113) (0.062) (0.061) (0.041) 0.201** 0.337*** -0.006 -0.141*** (0.072) (0.033) (0.043) (0.026) -0.420*** -0.007 0.264*** -0.150*** (0.085) (0.043) (0.048) (0.033)
Notes: The number of observations is 8819, 9505, 9887, 8876, 9776, and 8816 in Panels A through F respectively. All regressions include a constant and the full-set of controls, ZiW and ZiF . Tasks measures are from O*NET and vary by workers’ initial 4-digit occupations (ISCO). The availability of task measures across occupations determine the number of observations. Robust standard errors are 47 reported in parentheses. ∗ , ∗∗ and ∗∗∗ indicate significance at the 10 %, 5% and 1% levels respectively. Data Source: Statistics Denmark.
5.5
Technology, Trade, Man, Woman
We have shown that women tend to experience job polarization more strongly than men in Denmark. Controlling for worker and firm characteristics, women experience an additional gross churn of 0.7 years of employment away from mid-level and into high- and low-wage jobs (column 8 of Table 2). This finding is in line with the aggregate pattern in Denmark during the sample period (see Figure 2 in the Supplementary Appendix) and it also parallels results for the U.S. (Autor 2010, Figure 4). While job churn can entail substantial job switching costs, the fact that reductions in mid-level employment for women are accompanied by substantial gains in high-wage occupations may be considered evidence that women have more successfully adapted to shifts in demand that have eroded mid-level wage job opportunities than men (Autor 2010). At the same time to date we know very little on gender patterns of job polarization at the level of individual workers. In this section we begin by presenting such evidence for Denmark. First, is there reason to be more encouraged by the pattern for women than for men, as is the case in the U.S.? Second, we examine the role of trade for these genderspecific patterns of polarization. In particular, does import competition allow women to move up from mid-level to high-level jobs, or does import competition increase the chance that women fall down the job ladder to low-wage jobs? If women move both up and down the job ladder, are both shifts of equal importance quantitatively? The following section provides initial answers on these questions.
48
Table 9: Trade, Man, Woman Dep. Var. Panel A. Trade Trade*Female
JPiemp (1)
JPih,emp (2)
JPim,emp (3)
JPil,emp (4)
1.145*** (0.134) -0.575** (0.176)
0.364*** (0.065) -0.236** (0.082)
-0.546*** (0.078) 0.265* (0.104)
0.236*** (0.053) -0.074 (0.074)
1.112*** (0.135) -0.584*** (0.177) yes
0.317*** (0.064) -0.176* (0.081) yes
-0.549*** (0.079) 0.290** (0.104) yes
0.245*** (0.053) -0.117 (0.074) yes
1.110*** (0.140) -0.659*** (0.181) yes
0.283*** (0.065) -0.171* (0.082) yes
-0.562*** (0.082) 0.371*** (0.106) yes
0.266*** (0.055) -0.118 (0.075) yes
Panel B. Trade Trade*Female 2-digit ISCO Dummies Panel C. Trade Trade*Female 4-digit ISCO Dummies
Notes: In all regressions the number of observations is 10753. All regressions include a constant and the full-set of controls, ZiW and ZiF . Robust standard errors are reported in parentheses. ∗ , ∗∗ and ∗∗∗ indicate significance at the 10 %, 5% and 1% levels respectively. Data Source: Statistics Denmark.
To get a better feel of job polarization for women in Denmark, we begin by reporting results for a number of major occupations: managers, professionals, associate professionals and technicians, clerks, craft workers, and machine operators. We find that women that were technicians and clerks in year 1999 experience job polarization much less than men in these occupations, while women who were machine operators in 1999 experience job polarization to a far greater extent than male machine operators (Table A-11). Female machine operators in the textiles and clothing industry would typically be operating sewing machines. We also see that there is no difference between 1999 female and male managers in the extent to which they are employed in high-wage jobs during the subsequent years 2002 to 2009. In terms of education, we see that women with college education are less prone to job polarization than men, while in contrast women with at most high school education are more prone to job polarization than men (Table A-12, column 1, Panels A and C). 49
College-educated women are less likely than college-educated men to shift into high-wage occupations, while high school educated women are just as likely to move into high-wage jobs as men (column 2, Panels A and C). On the other hand, college educated women are less likely than college-educated men to move into low-wage jobs (column 4, Panel A). The fact that women with at most a high school diploma are substantially more prone to job polarization and they move more into low-wage jobs compared to men is in line with the general pattern observed in Denmark between 2000-2009 (see Supplementary Appendix, Figure 4). Next we examine employment polarization through trade in which women are separated from men (see Table 9). While job polarization overall has affected women more than men, the extent to which it is driven by trade is only about half as large for women as it is for men (column 1, Panel A). One might be concerned that this does not control for differences in the propensity with which men and women work in different occupations. To address this issue, we control in Panel B for two-digit and in Panel C for four-digit occupation fixed effects. The result remains largely unchanged: trade causes substantially less job polarization for women than for men. Before we look into the reasons for this it is important to ask whether trade-driven job polarization has had an impact on gender inequality. Columns 2, 3, and 4 of Table 9 show the impact of trade on cumulative employment in high-, mid-level, and low-wage occupations separately. The key finding here is that trade has shifted women less than men into high-wage jobs while at the same time trade has pushed women to the same extent as men into low-wage jobs.42 Trade is not a woman’s best friend because it raises gender inequality. An important question is why we estimate that trade affects women and men differently in leading to job polarization even when we control for detailed occupation characteristics (at the four-digit ISCO level). In the following we consider movements within and 42
This is true even if we look only at the point estimates and ignore the standard errors.
50
between industries along the lines of our earlier analysis in Table 6. In Table 10 we report employment changes for the three wage groups (high, mid-level, low) separately for manufacturing and non-manufacturing. The question is whether there are differences in how women and men behave in response to a trade shock across wage groups and sectors. We see that trade does not lower mid-level wage employment for women as much as for men within manufacturing (column 4). At the same time, trade does not shift women as much as men to high-wage jobs outside of manufacturing (column 3). This suggests that women stay more in the jobs that are hollowed out (mid-level manufacturing) and move less to high-paying jobs outside of manufacturing than men. Women tend to be more stayers, not movers, compared to men. This finding is limited, however, to job polarization caused by trade, as women in general are found experience job polarization more (they are more prone to leaving mid-wage jobs). While foregoing high-wage jobs outside of manufacturing relative to men (column 3), trade exposed womens’ tendency to stay in mid-level manufacturing jobs might benefit women because they do not shift as much as men into low-wage jobs within manufacturing as men (column 6).
If trade exposed women are less likely to be separated by their competition exposed firms compared to trade exposed men, then this may explain why women are relatively less prone to polarization caused by trade. To better understand this we focus on mid-wage level jobs within the manufacturing sector (see Table 11). In column (1) of Panel A is the cumulative employment in mid-wage occupations within the manufacturing sector. We then decompose the asymmetric effect of trade on women into the effect within and outside the initial firm in columns 2 and 3 respectively. We repeat the analysis with cumulative hours worked since women may be relatively likely to accept reduced hours of work at the initial firm instead of leaving for another job (Utar 2015). First, comparing the T radei coefficients of columns 1 and 2, in Panel A, Table 11 shows that the decrease in mid-level employment caused by trade is entirely due to employment 51
52
1.145*** (0.134) -0.575** (0.176) 0.966*** (0.125)
(1)
JP emp i
-0.003 (0.056) -0.090 (0.068) 0.090 (0.046)
Within Manuf. (2) 0.367*** (0.048) -0.146* (0.063) 0.216*** (0.041)
JP h,emp i Outside Manuf. (3) -0.661*** (0.074) 0.192* (0.096) -0.426*** (0.071)
Within Manuf. (4) 0.115** (0.044) 0.073 (0.059) -0.029 (0.041)
JP m,emp i Outside Manuf. (5)
0.127*** (0.038) -0.092* (0.045) -0.078* (0.031)
Within Manuf. (6)
0.109** (0.039) 0.018 (0.062) 0.283*** (0.043)
JP l,emp i Outside Manuf. (7)
Notes: In all columns the number of observations is 10753. A constant is included but not reported. All regressions include a constant and the full-set of controls, ZiW and ZiF . Robust standard errors are reported in parentheses. ∗ , ∗∗ and ∗∗∗ indicate significance at the 10 %, 5% and 1% levels respectively. Data Source: Statistics Denmark.
Female
Trade*Female
Trade
Dep. Var
Table 10: Trade, Man, Woman-Within and Outside Manufacturing
losses at the initial firm, not due to lower employment in other manufacturing firms. Given the importance of employment losses at the initial firm, does this differ between men and women? Table 11 shows that import shock does not have any significant asymmetric effect on women compared to men at the initial firm. That is, both women and men are affected by the shock to the same extent. This is reassuring because we have controlled for a substantial number of worker and firm characteristics (see Table 2). Panel A shows that 64 % of the asymmetric effect of trade on women (0.123/0.192) is due to mid-wage jobs within manufacturing but outside the initial firm. In terms of hours worked, 71 % of the effect is due mid-wage manufacturing jobs outside the initial firm. Thus, the difference in the extent that trade causes job polarization for men and women is largely due to gender differences in the propensity of moving to different sectors. What might explain these different responses in the face of a trade shock? It is important to keep in mind that women do not in general switch jobs and sectors less than men, because we have seen above that women are more likely than men to gain both highand low-wage employment in the services sector (Table 6). The question is why women are more attached to mid-level manufacturing jobs than men in the presence of a trade shock? We believe that this could be related to both the suddenness of the trade shock, as well as its massive size. The shock to labor demand through trade in the Danish textiles and clothing industry lead to a drastic resizing of the industry. When the industry is undergoing massive changes, the labor market environment might be similar to that in times of mass layoffs during recessions (Jacobson, Lalonde, and Sullivan 1993). In the more recent 2008 Great Recession, aggregate data for the U.S. suggests that women’s employment rates have recovered less well than those of men (Pew Research Center 2011). It is important to better understand similarities and differences between the labor market impact of trade on the one, and of recessions on the other hand. What we have documented is that a sudden and large-scale reduction in labor demand through trade is relatively more challenging for women than for men, compared to a gradual
53
shift of labor demand across sectors as, for example, the introduction of information technology and computers since the early 1980s.43
Table 11: Women and Mid-Wage Occupations Panel A. Dep. Var Trade Trade*Female Female Panel B. Dep. Var Trade Trade*Female Female
Cumulative employment Within Manuf. (1) -0.661*** (0.074) 0.192* (0.096) -0.426*** (0.071)
in mid-wage occupations within manufacturing 2002-2009 Within the Initial Firm Outside the Initial Firm (2) (3) -0.668*** 0.007 (0.058) (0.059) 0.069 0.123 (0.076) (0.074) -0.052 -0.375*** (0.061) (0.052)
Cumulative hours worked in mid-wage occupations within manufacturing 2002-2009 Overall Within Manuf. Within the Initial Firm Outside the Initial Firm (1) (2) (3) -0.768*** -0.688*** -0.080 (0.085) (0.061) (0.071) 0.205 0.059 0.146 (0.116) (0.082) (0.096) -0.390*** -0.038 -0.352*** (0.091) (0.066) (0.075)
Notes: In all columns the number of observations is 10753. A constant is included but not reported. All regressions include a constant and the full-set of controls, ZiW and ZiF . Robust standard errors are reported in parentheses. ∗ ∗∗ , and ∗∗∗ indicate significance at the 10 %, 5% and 1% levels respectively. Data Source: Statistics Denmark.
6
Conclusions
Using administrative employee-employer matched data from Statistics Denmark covering all residents of Denmark, we document that Denmark, like many other rich countries, experienced job polarization in recent years which coincided with a dramatic increase in Chinese exports to these countries. We document that employment share changes of workers in the manufacturing sector exhibit a U-shaped pattern away from mid43
We have also examined how gender affected the impact of technology factors on job polarization by employing the interaction of F emalei and O N ETi , see Table A-13. Overall, these results show that the gender difference plays an important role in our finding that trade and technology factors are distinct causes of job polarization in Denmark.
54
level and toward high- and low-wage jobs while workers in the service sector experience employment changes that are increasing with wages and skills between 2000 and 2009. This is important for at least two reasons. The trade effect is relatively concentrated in manufacturing, whereas technical change affects also, and increasingly, sectors outside of manufacturing. Second, while technology has gradually changed the workplace in industrialized countries since the early 1980s, the intensification of import competition was felt more suddenly, especially in the early 2000s. To investigate the causal link between Chinese imports and job polarization, we exploit the exogenous abolishment of trade quotas for China associated with her entry into the WTO, and utilize heterogeneity in workers’ exposure to this trade shock within the same narrow industry. This allows us to distinguish the effects of trade from technical change, whose effect on job polarization is well established. We have shown that trade can explain the U-shaped changes in employment that are the hallmark of job polarization. Trade leads to job polarization mainly by pushing workers from initially abundant midlevel jobs in manufacturing towards both high- and low-wage jobs in services. We have seen that trade affects most strongly manual labor, workers performing both routine and non-routine tasks. We find that the shift from manufacturing to services is crucial for this trade effect. Worker level evidence sharpens the focus because it lets us observe, in the data, individual worker transitions that are the link between sectors. It may be that worker-level information is one of the reasons why we have found more evidence that trade causes job polarization than other studies before us. We also find that while women experience overall job polarization more than men, women contribute less than men to the part of job polarization that is related to trade. One reason for the former may be that even within fine occupation classes women might tend to perform more routine tasks than men, and it appears that women adjust well in terms job changes in response to gradual changes. In contrast, in the face of the
55
sudden reduction in labor demand caused by trade, women do not adjust as well as men. While it is clear that a sudden shock does not give as much opportunity to plan and prepare, for example re-training, what explains the gender difference? One reason might be that because women tend to be more strongly involved with child care and other joint household activities than men, women have higher short-term adjustment costs. It could also be that the on average lower labor market attachment of women compared to men makes it harder for women to move when the change is sudden as opposed to gradual. The mid-level wage jobs that job polarization feeds on are primarily manufacturing jobs. Can one make a case that this matters? A shrinking manufacturing sector in industrialized countries might be quite different from other instances of structural change; it may have broader implications, by affecting industry clustering, city formation, or innovation.44 Mid-level jobs might have important positive spillovers. To the extent that trade causes job polarization, it determines the size of these spillovers, and trade should be considered in the formulation of policies that are designed to internalize these spillovers. What are the implications of job polarization on inequality? While our results show that trade tends to increase gender inequality because it shifts men to a greater extent into high-wage jobs than women, this may be all but a small part of the impact of trade on inequality. As international economic integration increases, if imports from poor countries continue to hollow out the middle of the wage distribution in rich countries, the implied increase in inequality might affect social cohesion and broader political outcomes in rich countries, and beyond. 44
On clustering and city formation, see Moretti (2012). The importance of near-by manufacturing for innovation is also an important question (Keller, Yeaple, and Zolas 2015).
56
References [1] Acemoˇ glu, Daron 2002. “Technical Change Inequality, and the Labor Market”, Journal of Economic Literature, 40(1) : 7-72. [2] Acemoˇ glu, Daron, and David Autor 2010. “Skills, Tasks and Technologies: Implications for Employment and Earnings”, Handbook of Labor Economics Volume 4, Orley Ashenfelter and David E. Card (eds.), Amsterdam: Elsevier, forthcoming. [3] Autor, David 2010. “The Polarization of Job Opportunities in the U.S. Labor Market: Implications for Employment and Earnings”, Center for American Progress and The Hamilton Project. [4] Autor, David 2013. “The “task approach” to labor markets: an overview”, Journal for Labour Market Research, 46(3): 185-199. [5] Autor, David, and David Dorn 2013. “The Growth of Low-Skill Service Jobs and the Polarization of the US Labor Market”, American Economic Review, 103(5): 1553-1597. [6] Autor, David, David Dorn, and Gordon Hanson. 2013. “The China Syndrome: Local Labor Market Effects of Import Competition in the United States”, American Economic Review, 103(6): 2121-2168. [7] Autor, David, David Dorn, and Gordon Hanson. 2014. “Untangling Trade and Technology: Evidence from local Labor Markets”, Economic Journal, forthcoming. [8] Autor, David, David Dorn, Gordon Hanson and Jae Song. 2014. “Trade Adjustment: Worker Level Evidence”, The Quarterly Journal of Economics, 129: 1799-1860. [9] Autor, David, Lawrance F. Katz, and Melissa S. Kearney 2006. “The Polarization of the U.S. Labor Market”, American Economic Review: Papers and Proceedings, 96(2), 189-193. [10] Autor, David, Lawrance F. Katz, and Melissa S. Kearney 2008. “Trends in U.S. Wage Inequality: Revising the Revisionists”, The Review of Economics and 57
Statistics, 90(2), 300-323. [11] Autor, David, Levy, Frank, and Richard Murnane 2003. “The Skill-Content of Recent Technological Change: An Empirical Investigation”, Quarterly Journal of Economics, 118, 1279-1333. [12] Bernard, Andrew B., Bradford J. Jensen and Peter K. Schott 2006. “Survival of the best fit: Exposure to low-wage countries and the (uneven) growth of U.S. manufacturing plants”, Journal of International Economics, Elsevier, vol. 68(1), 219-237. [13] Blinder, Alan, and Alan Krueger 2013. “Alternative Measures of Offshorability: A Survey Approach”, Journal of Labor Economics, 31(2): S97-S128 [14] Bloom, Nicholas, Mirko Draca, and John Van Reenen 2012. “Trade induced technical change? The impact of Chinese imports on innovation and information technology”, NBER Working Paper, No. 16717. [15] Brambilla, Irene, Amit Khandelwal and Peter Schott 2010. “China’s Experience Under the Multifiber Arrangement (MFA) and the Agreement on Textile and Clothing (ATC)”, Robert Feenstra and Shang-Jin Wei (Eds), China’s Growing Role in World Trade, NBER, Cambridge: MA. [16] Bunzel, Henning 2008. “The LMDG Data Sets”, mimeo, Univeristy of Aarhus. [17] Dustmann, Christian, Johannes Ludsteck, and Uta Schonberg 2009. “Revisiting the German Wage Structure”, The Quarterly Journal of Economics, 124, 843-881. [18] Feenstra, Robert 2000. Introduction to “The Impact of International Trade on Wages”, In Robert C. Feenstra, ed., The Impact of International Trade on Wages, Chicago, University of Chicago Press, 1-11. [19] Firpo, Sergio, Nicole Fortin, and Thomas Lemieux 2011. “Occupational Tasks and Changes in the Wage Structure”, IZA Discussion Paper 5542, INstitute for the Study of Labor (IZA).
58
[20] Goos, Maarten, and Alan Manning 2007. “Lousy and Lovely Jobs: The Rising Polarization of Work in Britain”, The Review of Economics and Statistics, 89(1), 118-133. [21] Goos, Maarten, Alan Manning, and Anna Salomons 2009. “Job Polarization in Europe”, American Economic Review: Papers and Proceedings, 99(2), 58-63. [22] Goos, Maarten, Alan Manning, and Anna Salomons 2014. “Explaining Job Polarization: Routine-Biased Technological Change and Offshoring”, American Economic Review, 104(8), 2509-2526. [23] Groes, Fane, Philipp Kircher, and Iorii Manovskii 2015. “The U-Shapes of Occupational Mobility”, Review of Economic Studies, 82(2). [24] Grossman, Gene and Esteban Rossi-Hansberg 2008. “Trading Tasks: A Simple Theory of Offshoring”, American Economic Review, 98:5, 19781997. [25] Harrigan, James, Ariell Reshef, and Farid Toubal 2015. “The March of the Techies: Technology, Trade, and Job Polarization in France, 1994-2007”, presented at Barcelona GSE Summer Forum 2015. [26] Hummels, David, Rasmus Jorgensen, Jacob Munch, and Chong Xiang 2014. “The Wage Effects of Offshoring: Evidence From Danish Matched WorkerFirm Data”, American Economic Review, 104(6): 1597-1629. [27] Jacobson, Louis, Robert LaLonde and Daniel Sullivan 1993. “Earnings Losses of Displaced Workers”, American Economic Review, 83(4): 685-709. [28] Keller, Wolfgang, Stephen Yeaple, and Nick Zolas 2015. “Innovation in the Age of Offshoring”, in progress. [29] Khandelwal, Amit, Peter Schott and Shang-Jin Wei 2013. “Trade Liberalization and Embedded Institutional Reform: Evidence from Chinese Exporters”, American Economic Review, Vol. 103(6): 2169-2195. [30] Michaels, Guy, Ashwini Natraj and John Van Reenen 2014. “Has ICT Polarized Skill Demand? Evidence From Eleven Countries Over Twenty-Five Years”, The Review of Economics and Statistics, 96(1), 60-77.
59
[31] Moretti, Enrico 2012. The New Geography of Jobs. New York: Houghton Mifflin Harcourt [32] Pew Research Center 2011. “Two Years of Economic Recovery: Women Lose Jobs, Men Find Them”, by Rakesh Kochhar, July 6, 2011. [33] Pierce, Justin R. and Peter K. Schott. 2014. “The Surprisingly Swift Decline of U.S. Manufacturing Employment”, NBER Working Paper No. 18655. [34] Spitz-Oener, Alexandra 2006. “Technical Change, Job Tasks and Rising Educational Demand: Looking Outside the Wage Structure”, Journal of Labor Economics, 24, 235-270. [35] Timmermans, Bram 2010. “The Danish Integrated Database for Labor Market Research: Towards Demystification for the English Speaking Audience”, DRUID Working Papers 10-16. [36] Utar, Hale 2014. “When the Floodgates Open: ”Northern” Firms’ Response to Removal of Trade Quotas on Chinese Goods”, American Economic Journal: Applied Economics, 6(4): 226-250. [37] Utar, Hale 2015. “Workers beneath the Floodgates: Impact of Low-Wage Import Competition and Workers’ Adjustment”, Bielefeld Working Papers in Economics and Management, No.12. [38] Utar, Hale, and Luis Torres Ruiz. 2013. “International Competition and Industrial Evolution: Evidence from the Impact of Chinese Competition on Mexican Maquiladoras”, Journal of Development Economics, 105: 267-287.
Appendix A
60
s
an
ni ci
ch
Te
an
IS
IS C
0
Density
O -1 C d As O-2 M an so P a IS r ci at ofe ger C O e s s -6 Pr sio of n Sk I es als ille SC I d O-5 SC sion Ag O al IS s r a Ser -4 C O nd vic Cle -8 e rk F W Pl s is IS h o a IS nt CO ery rke an -7 C W rs O -9 d M Cra or ke El f a em ch t W rs en ine ork er O ta p s ry O era cc to IS up rs C at O i -3 U ons Te nk ch no ni w ci n an IS C s O an IS -1 C d As O-2 M a so Pr na IS ci at ofe ger C O e s s -6 Pr sio of n Sk I es als ille SC I d O-5 SC sion Ag O al IS s r a Ser -4 C O nd vic Cle -8 e rk F W Pl s is I h o a S IS nt CO ery rke an -7 C W rs O d C or -9 r M ke El a af em ch t W rs en ine ork er O ta p s ry O era cc to up rs at i U ons nk no w n
-3
O
C
IS
.1
.2
.3 .4 -3
C O
IS
(a) All Workers
Workers employed in exposed firms
OneDigOc
61 Workers employed in non-exposed firms
(b) By Import Competition
FigureGraphs A-1:by AffW Histogram of Workers across Major Occupations in 1999 OneDigOc
IS
IS
-9
-5
O
ar y
O
U nk
no w
n
ns
up at io
rs
rs
or ke
rs
or ke
pe ra to
O
cc
e
W
W or ke rs
rk s
io na ls C le
tW
ra f
C
ry
he
ac hi n
M
-7
O
Fi s
-4 ic e
O
C Se rv
en t
em
El
an d
IS C
Ag ra nd
an t
Pl
C O
-8
d
C
IS
on al s
si es s
Pr of
es
er s
ag
M an
Pr of
-1
O
-2
C
at e
ci
so
O
C
IS As
IS
d
an
ille
Sk
O
C
IS
-6
C O
s
an
ni ci
ch
Te
IS
0
.1
Density .3 .2
.4
Table A-1: Workers’ Characteristics in 1999 by Selective Occupation Groups ISCO 1: Managers Age Female Immigrant College Educated Vocational School Educated At most High School ISCO 2: Professionals Age Female Immigrant College Educated Vocational School Educated At most High School ISCO 3: Technicians and Associate Professionals Age Female Immigrant College Educated Vocational School Educated At most High School ISCO 4: Clerks Age Female Immigrant College Educated Vocational School Educated At most High School ISCO 7: Craft Workers Age Female Immigrant College Educated Vocational School Educated At most High School ISCO 8: Plant and Machine Operators Age Female Immigrant College Educated Vocational School Educated At most High School ISCO 9: Elementary Occupations Age Female Immigrant College Educated Vocational School Educated At most High School Notes: Data Source: Statistics Denmark.
62
Mean
SD
N
44.912 0.211 0.020 0.256 0.436 0.297
8.302 0.408 0.140 0.437 0.496 0.457
555 555 555 555 555 555
39.461 0.362 0.013 0.612 0.178 0.184
9.634 0.482 0.114 0.489 0.383 0.389
152 152 152 152 152 152
37.756 0.652 0.028 0.371 0.365 0.226
9.097 0.477 0.164 0.483 0.482 0.418
1338 1338 1338 1338 1338 1338
36.423 0.780 0.019 0.151 0.531 0.317
9.569 0.414 0.137 0.358 0.499 0.466
1093 1093 1093 1093 1093 1093
40.702 0.377 0.040 0.061 0.487 0.425
10.381 0.485 0.197 0.240 0.500 0.495
916 916 916 916 916 916
40.617 0.594 0.090 0.032 0.276 0.665
10.181 0.491 0.287 0.175 0.447 0.472
4679 4679 4679 4679 4679 4679
38.046 0.510 0.052 0.027 0.325 0.627
11.247 0.500 0.222 0.162 0.469 0.484
923 923 923 923 923 923
Table A-2: Workers’ Characteristics in 1999 by Selective Occupations and Import Exposure Mean SD N Weaving and Knitting Machine Operators employed in non-exposed firms Age 40.566 10.479 304 Female 0.247 0.432 304 Immigrant 0.089 0.285 304 College Educated 0.043 0.203 304 Years of Experience in the Labor Market 15.740 5.439 304 Unemployment History Index 1019.102 1509.095 304 Weaving and Knitting Machine Operators employed in exposed firms Age 41.966 9.459 205 Female 0.288 0.454 205 Immigrant 0.112 0.316 205 College Educated 0.049 0.216 205 Years of Experience in the Labor Market 16.493 4.813 205 Unemployment History Index 1122.215 1557.671 205 Sewing Machine Operators employed in non-exposed firms Age 42.849 9.368 537 Female 0.946 0.226 537 Immigrant 0.143 0.351 537 College Educated 0.020 0.142 537 Years of Experience in the Labor Market 15.294 5.579 537 Unemployment History Index 2573.940 2511.583 537 Sewing Machine Operators employed in exposed firms Age 43.257 10.026 915 Female 0.957 0.202 915 Immigrant 0.072 0.259 915 College Educated 0.026 0.160 915 Years of Experience in the Labor Market 16.216 5.255 915 Unemployment History Index 1429.028 1695.268 915 Post-Processing Textile Machine Operators employed in non-exposed firms Age 40.478 9.894 502 Female 0.329 0.470 502 Immigrant 0.084 0.277 502 College Educated 0.064 0.245 502 Years of Experience in the Labor Market 14.964 5.535 502 Unemployment History Index 1374.861 1736.771 502 Post-Processing Textile Machine Operators employed in exposed firms Age 39.030 9.667 133 Female 0.323 0.470 133 Immigrant 0.045 0.208 133 College Educated 0.030 0.171 133 Years of Experience in the Labor Market 15.609 5.145 133 Unemployment History Index 1464.586 2002.921 133 Cutting, Rope Making, Netting, Leather Textile Machine Operators in non-exposed firms Age 39.888 9.311 251 Female 0.199 0.400 251 Immigrant 0.008 0.089 251 College Educated 0.020 0.140 251 Years of Experience in the Labor Market 16.275 4.791 251 Unemployment History Index 1713.175 2117.432 251 Cutting, Rope Making, Netting, Leather Text. Machine Operators in exposed firms Age 40.466 10.201 221 Female 0.674 0.470 221 Immigrant 0.081 0.274 221 College Educated 0.032 0.176 221 Years of Experience in the Labor Market 15.330 5.396 221 Unemployment History Index 1618.543 2068.907 221 Notes: Data Source: Statistics Denmark.
63
The T&C Workers' Industry Affiliation in 2009
43%
52%
Agr, Fishing, Mining Manufacturing Service
Construction Others
Figure A-2: Industry Affiliation of the 1999 T&C Workers in 2009 (conditional on working)
The T&C Workers' Industry Affiliation in 2009 Exposed
Non-Exposed
34%
44% 50%
60%
Agr, Fishing, Mining Manufacturing Service
Construction Others
Graphs by ImportCompetition
Figure A-3: Industry Affiliation of the 1999 T&C Workers in 2009 by exposure to import competition (conditional on working)
64
65
5.460 5.320 5.101 4.945 5.013 4.898 4.873 4.799
5.032 5.000 4.907 4.842 4.842
Mean Log Hourly Wage in 1991
5.430 5.269 5.078
Median Log Hourly Wage in 1991
0.154
0.106
0.149 0.065 0.170
0.042 0.128 0.186
Employment Share in 1991
5
9 (except 92)
7 8 4
1 2 3
Corresponding Major ISCO
Data are from 1991. Values are expressed in 2000 Danish Kroner. All hourly wages are calculated among workers with full-time jobs employed continuously with at least one year tenure. Employment shares are calculated using the number of employees and excluding army and agriculture/fishery occupations. Data Source: Statistics Denmark.
High-Wage Legislators, Senior Officials, and Managers Professionals Technicians and Associate Professionals Mid-Wage Craft and Related Trade Workers Plant and Machine Operators and Assemblers Clerks Low-Wage Elementary Occupations in Sales, Services, Mining, Construction, Manufacturing, and Transport Service Workers and Shop Sales Workers
Occupation
Table A-3: Ranking of Major Occupations in Denmark
Appendix B
Table A-4: Job Polarization and Import Shock Dep. Var. Trade
ln(9 + JPiemp ) (1) 0.132*** (0.015) [0.079]
ln(1 + JPih,emp ) (2) 0.084*** (0.012) [0.053]
ln(1 + JPim,emp ) (3) -0.109*** (0.015) [-0.064]
ln(1 + JPil,emp ) (4) 0.073*** (0.013) [0.054]
Notes: In all regressions the number of observations is 10753. All regressions include a constant and the full-set of controls, ZiW and ZiF . Robust standard errors are reported in parentheses. ∗ , ∗∗ and ∗∗∗ indicate significance at the 10 %, 5% and 1% levels respectively. Standardized coefficients are reported in square brackets. Data Source: Statistics Denmark.
Table A-5: Job Polarization and Import Shock–Alternative Trade Exposure Dep. Var. T^ rade
JPiemp (1) 3.220*** (0.316) [0.092]
JPih,emp (2) 0.904*** (0.146) [0.051]
JPim,emp (3) -1.582*** (0.181) [-0.079]
JPil,emp (4) 0.734*** (0.138) [0.052]
Notes: T^ rade is the revenue share of the quotas at worker i’s initial employer in 1999. In all regressions the number of observations is 10753. All regressions include a constant and the full-set of controls, ZiW and ZiF . Robust standard errors are reported in parentheses. ∗ , ∗∗ and ∗∗∗ indicate significance at the 10 %, 5% and 1% levels respectively. Standardized coefficients are reported in square brackets. Data Source: Statistics Denmark.
66
Table A-6: Polarization and Trade by Initial Occupation Dep. Var. Panel A. Trade Trade*Managers
JPiemp JPih,emp (1) (2) Managers (ISCO: 1)
JPim,emp (3)
JPil,emp (4)
0.830*** (0.093) -0.972** (0.321)
-0.404*** (0.055) 0.393** (0.140)
0.186*** (0.040) -0.149 (0.081)
-0.398*** (0.053) 1.088*** (0.322)
0.184*** (0.038) -0.431 (0.233)
0.240*** (0.040) -0.429 (0.241)
Panel B.
Professionals (ISCO: 2)
Trade
0.810*** (0.090) -2.349** (0.724)
Trade*Profs
0.228*** (0.040) -0.830 (0.496)
Panel C.
Associate Professionals and Technicians (ISCO: 3)
Trade
0.864*** (0.096) -0.706** (0.259)
Trade*Technicians
0.191*** (0.039) 0.203 (0.169)
Panel D.
Clerks (ISCO: 4)
Trade
0.879*** (0.093) -0.977** (0.313)
Trade*Clerks
0.258*** (0.041) -0.396** (0.150)
Panel E.
Craft Workers (ISCO: 7)
Trade
0.803*** (0.093) -0.328 (0.321)
Trade*Craftsmen
0.214*** (0.042) 0.029 (0.130)
-0.438*** (0.057) 0.449*** (0.126)
0.235*** (0.042) -0.460*** (0.077)
-0.442*** (0.054) 0.570** (0.184)
0.179*** (0.041) -0.010 (0.087)
-0.402*** (0.054) 0.241 (0.204)
0.187*** (0.040) -0.116 (0.114)
Panel F.
Plant and Machine Operators and Assemblers (ISCO: 8)
Trade
0.360** (0.120) 0.937*** (0.175)
Trade*Operators
0.252*** (0.064) -0.081 (0.073)
-0.080 (0.065) -0.681*** (0.104)
0.028 (0.046) 0.337*** (0.077)
Notes: In all regressions the number of observations is 10753. All regressions include a constant and the full-set of controls, ZiW and ZiF as well as the 2-digit ISCO (initial occupation) dummies. Robust standard errors are reported in parentheses. ∗ , ∗∗ and ∗∗∗ indicate significance at the 10 %, 5% and 1% levels respectively.
67
Table A-7: Trade induced Job Polarization among Machine Operators Dep. Var.
JPiemp JPih,emp JPim,emp (1) (2) (3) Among Machine Operators (ISCO: 82)
JPil,emp (4)
Trade
1.079*** (0.148)
0.323*** (0.072)
0.097* (0.039)
-0.659*** (0.095)
Notes: Estimation sample consist of workers who were machine operators (ISCO=82) in the textile and clothing sector in 1999. In all regressions the number of observations is 3773. All regressions include a constant and the full-set of controls, ZiW and ZiF . Robust standard errors are reported in parentheses. ∗ ∗∗ , and ∗∗∗ indicate significance at the 10 %, 5% and 1% levels respectively.
Table A-8: Polarization and Trade by Education Dep. Var. Panel A. Trade Trade*College
JPiemp JPih,emp (1) (2) College Educated
JPim,emp (3)
JPil,emp (4)
0.931*** (0.095) -0.998*** (0.270)
-0.464*** (0.056) 0.609*** (0.140)
0.214*** (0.042) -0.178* (0.078)
-0.388*** (0.063) -0.016 (0.106)
0.250*** (0.048) -0.159* (0.075)
-0.279*** (0.071) -0.226* (0.100)
0.091 (0.049) 0.201** (0.073)
0.253*** (0.040) -0.212 (0.163)
Panel B.
Vocational School
Trade
0.852*** (0.108) -0.103 (0.180)
Trade*Vocational
0.214*** (0.048) 0.040 (0.082)
Panel C.
at most High School Diploma
Trade
0.572*** (0.126) 0.480** (0.170)
Trade*HS
0.201** (0.064) 0.054 (0.077)
Notes: In all regressions the number of observations is 10753. All regressions include a constant and the full-set of controls, ZiW and ZiF . Robust standard errors are reported in parentheses. ∗ , ∗∗ and ∗∗∗ indicate significance at the 10 %, 5% and 1% levels respectively.
68
69 no
0.912*** (0.132) 0.168** (0.058)
no
0.922*** (0.132) 0.142* (0.064) -0.104 (0.097)
JPihrs (2)
no
0.910*** (0.132) 0.139* (0.064) -0.339** (0.128) 0.417** (0.156)
JPihrs (3)
no
0.514** (0.177) 0.032 (0.072) -0.329* (0.128) 0.434** (0.157) 0.270*** (0.076)
JPihrs (4)
no
-0.737*** (0.134)
0.774*** (0.135) 0.118 (0.064) -0.003 (0.139) 0.454** (0.156)
JPihrs (5)
-0.088 (0.161) no
0.962*** (0.154) 0.138* (0.063) -0.334** (0.129) 0.455** (0.166)
JPihrs (6)
yes
0.864*** (0.138)
JPihrs (7)
Notes: In all columns the number of observations is 8702. A constant and the full set of controls, ZiW and ZiF , are included in all regressions but not reported. “Offshoring” is the offshorability index of the corresponding two digit initial ISCO-88 occupation code of a worker. It is constructed by Goos, Manning and Salomons (2014). “ICT” is the index variable indicating the degree in which the initial (4-digit) occupation of a worker interacts with computers. It is an O*NET (version 14) variable and varies across 4-digit occupations. “RTI” is the routine intensity index of the corresponding two digit initial ISCO-88 occupation code of a worker. The RTI index follows Autor, Levy and Murnane (2003) and Autor and Dorn (2012). Control variables are described in the notes of Table 3. Robust standard errors are reported in parentheses. ∗ , ∗∗ and ∗∗∗ indicate significance at the 10 %, 5% and 1% levels respectively. Data Source: Statistics Denmark.
4-Digit ISCO Dummies
Trade*RTI
Trade*ICT
Trade*Offshoring
RTI
ICT
Offshoring
Trade
JPihrs (1)
Table A-9: Job Polarization, Offshoring, Technology, and Trade - Hours Worked
70 no
0.992*** (0.177) 0.186* (0.077)
no
1.000*** (0.175) 0.164 (0.087) -0.088 (0.132)
JPiinc (2)
no
0.988*** (0.175) 0.161 (0.087) -0.336 (0.174) 0.438* (0.217)
JPiinc (3)
no
0.764** (0.243) 0.101 (0.097) -0.330 (0.174) 0.447* (0.217) 0.153 (0.095)
JPiinc (4)
ZiW
no
-0.738*** (0.176)
0.851*** (0.177) 0.140 (0.086) 0.002 (0.187) 0.475* (0.217)
JPiinc (5)
ZiF ,
-0.114 (0.211) no
1.055*** (0.207) 0.160 (0.086) -0.329 (0.175) 0.487* (0.229)
JPiinc (6)
yes
0.951*** (0.180)
JPiinc (7)
Notes: In all columns the number of observations is 8702. A constant and the full set of controls, and are included in all regressions but not reported. “Offshoring” is the offshorability index of the corresponding two digit initial ISCO-88 occupation code of a worker. It is constructed by Goos, Manning and Salomons (2014). “ICT” is the index variable indicating the degree in which the initial (4-digit) occupation of a worker interacts with computers. It is an O*NET (version 14) variable and varies across 4-digit occupations. “RTI” is the routine intensity index of the corresponding two digit initial ISCO-88 occupation code of a worker. The RTI index follows Autor, Levy and Murnane (2003) and Autor and Dorn (2012). Control variables are described in the notes of Table 3. Robust standard errors are reported in parentheses. ∗ , ∗∗ and ∗∗∗ indicate significance at the 10 %, 5% and 1% levels respectively. Data Source: Statistics Denmark.
4-Digit ISCO Dummies
Trade*RTI
Trade*ICT
Trade*Offshoring
RTI
ICT
Offshoring
Trade
JPiinc (1)
Table A-10: Job Polarization, Offshoring, Technology, and Trade - Earnings
Table A-11: Polarization and Women by Initial Occupation Dep. Var. Panel A. Trade Managers*Female
JPiemp (1) Managers
JPih,emp (2)
JPim,emp (3)
JPil,emp (4)
0.778*** (0.090) -0.575 (0.385)
0.216*** (0.040) -0.082 (0.292)
-0.383*** (0.052) 0.304 (0.164)
0.178*** (0.038) -0.189 (0.126)
0.216*** (0.040) 0.456 (0.458)
-0.383*** (0.052) 0.805* (0.360)
0.179*** (0.038) -1.043*** (0.207)
Panel B.
Professionals
Trade
0.778*** (0.090) -1.392 (0.742)
Profs*Female Panel C.
Associate Professionals and Technicians
Trade
0.775*** (0.089) -2.174*** (0.261)
Technicians*Female Panel D.
Clerks
Trade
0.734*** (0.090) -2.769*** (0.360)
Clerks*Female Panel E.
Craft Workers
Trade
0.775*** (0.090) 0.635 (0.368)
Craft*Female
0.216*** (0.040) -0.640*** (0.175)
-0.382*** (0.052) 1.104*** (0.125)
0.178*** (0.038) -0.430*** (0.075)
0.209*** (0.040) -0.472* (0.184)
-0.357*** (0.052) 1.659*** (0.208)
0.168*** (0.038) -0.638*** (0.109)
0.216*** (0.040) 0.202 (0.150)
-0.381*** (0.052) -0.490* (0.234)
0.178*** (0.038) -0.057 (0.138)
-0.334*** (0.052) -1.089*** (0.111)
0.156*** (0.038) 0.490*** (0.078)
Panel F.
Machine Operators
Trade
0.684*** (0.090) 2.073*** (0.184)
Operators*Female
0.194*** (0.040) 0.494*** (0.078)
Notes: In all regressions the number of observations is 10753. All regressions include a constant and the full-set of controls, ZiW and ZiF as well as the 2-digit ISCO (initial occupation) dummies. Robust standard errors are reported in parentheses. ∗ , ∗∗ and ∗∗∗ indicate significance at the 10 %, 5% and 1% levels respectively.
71
Table A-12: Polarization and Women by Education Dep. Var. Panel A. Trade College*Female
JPiemp JPih,emp (1) (2) College Educated
JPim,emp (3)
JPil,emp (4)
0.816*** (0.089) -1.613*** (0.269)
-0.394*** (0.052) 0.786*** (0.140)
0.194*** (0.038) -0.452*** (0.078)
-0.391*** (0.052) 0.190 (0.106)
0.190*** (0.038) -0.204** (0.073)
-0.387*** (0.052) -0.392*** (0.102)
0.187*** (0.038) 0.397*** (0.071)
0.229*** (0.040) -0.375* (0.165)
Panel B.
Vocational School
Trade
0.810*** (0.089) -0.371* (0.180)
Vocational*Female
0.229*** (0.040) 0.023 (0.082)
Panel C.
at most High School Diploma
Trade
0.802*** (0.089) 0.844*** (0.172)
at most High School*Female
0.228*** (0.040) 0.056 (0.078)
Notes: In all regressions the number of observations is 10753. All regressions include a constant and the full-set of controls, ZiW and ZiF . Robust standard errors are reported in parentheses. ∗ , ∗∗ and ∗∗∗ indicate significance at the 10 %, 5% and 1% levels respectively.
72
Table A-13: Tasks, Trade, Man, Woman Dep. Var. Panel A. Trade Repetitive Motions Female*Repetitive Motions Panel B. Trade GBC Female*GBC Panel C. Trade Evaluating Female*Evaluating Panel D. Trade Thinking Female*Thinking Panel E. Trade ICT Female*ICT Panel F. Trade Interpersonal Female*Interpersonal
JPiemp JPih,emp JPim,emp JPil,emp (1) (2) (3) (4) Routine Manual Tasks Tasks with Repetitive Motions 0.752*** 0.246*** -0.368*** 0.138** (0.099) (0.044) (0.057) (0.043) -0.550*** -0.405*** 0.169** 0.024 (0.095) (0.049) (0.055) (0.032) 0.884*** 0.139* -0.431*** 0.313*** (0.104) (0.055) (0.059) (0.038) Non-Routine Manual Tasks Tasks with Gross Body Coordination (GBC) 0.800*** 0.195*** -0.405*** 0.200*** (0.094) (0.041) (0.055) (0.041) -0.854*** -0.611*** 0.355*** 0.112*** (0.103) (0.059) (0.055) (0.034) 1.343*** 0.241*** -0.773*** 0.329*** (0.119) (0.065) (0.064) (0.043) Routine Cognitive Tasks Evaluating Information to Determine Compliance with Standards 0.807*** 0.199*** -0.409*** 0.199*** (0.093) (0.042) (0.054) (0.040) 0.009 -0.105* 0.013 0.127*** (0.083) (0.041) (0.049) (0.030) -0.004 0.104 -0.033 -0.141** (0.116) (0.058) (0.066) (0.046) Abstract Tasks Tasks that involves thinking creatively (Thinking) 0.796*** 0.226*** -0.396*** 0.174*** (0.098) (0.044) (0.057) (0.043) 0.377*** 0.265*** -0.089 0.023 (0.096) (0.047) (0.056) (0.034) -0.488*** -0.187*** 0.135* -0.165*** (0.104) (0.056) (0.058) (0.040) ICT Tasks that involves interacting with computers (ICT) 0.754*** 0.171*** -0.385*** 0.198*** (0.093) (0.041) (0.054) (0.040) 0.744*** 0.512*** -0.270*** -0.038 (0.092) (0.048) (0.052) (0.032) -1.308*** -0.242*** 0.715*** -0.351*** (0.102) (0.052) (0.057) (0.038) Face to Face Communication Intensive Tasks Tasks that involve establishing interpersonal relationships 0.678*** 0.216*** -0.329*** 0.132** (0.099) (0.044) (0.058) (0.042) 0.534*** 0.420*** -0.147*** -0.033 (0.075) (0.039) (0.043) (0.025) -0.961*** -0.151*** 0.486*** -0.324*** (0.087) (0.045) (0.049) (0.033)
Notes: The number of observations is 8819, 9505, 9887, 8876, 9776, and 8816 in Panels A through F respectively. All regressions include a constant and the full-set of controls, ZiW and ZiF . Tasks measures are from O*NET and vary by workers’ initial 4-digit occupations (ISCO). The availability of task measures across occupations determine the number of observations. Robust standard errors are reported in parentheses. ∗ , ∗∗ and ∗∗∗ indicate significance at the 10 %, 5% and 1% levels respectively. 73 Data Source: Statistics Denmark.
