Job polarization in domestic labor markets and its drivers - An empirical investigation

Master Thesis

Job polarization in domestic labor markets and its drivers -

An empirical investigation

June 18, 2019

DD MSc Int. Economics & Business/ MA International Economics

Supervisor:

Assc. Prof. Dr. Gaaitzen J. de Vries, University of Groningen

Co-Assessor:

ii

Abstract

This paper studies the occurrence of job polarization for a sample of 89 advanced, emerging and developing countries, by analyzing changes in occupational structures in the two decades 1990-2010. The relationship between changes in domestic employment distributions and the potential drivers technological change and job offshoring, as well as reinforcing factors to these forces is investigated by means of cross sectional regression models. Job polarization appears to occur increasingly with economic development, as it is pervasive across advanced economies, less pronounced in emerging economies, and not observable in developing country labor markets. Routine-biased technological change is found to be the predominant driver of the phenomenon, strongly affecting advanced economies and more moderately emerging economies, while job relocation is not found to play a significant role.

Keywords: Job Polarization, Routine-biased technological change,

iii

Content

List of abbreviations ... iv

1 Introduction ... 1

2 Theoretical background and hypotheses development... 3

2.1 Job polarization ... 3

2.2 A task framework for the analysis of labor market polarization ... 5

2.3 Drivers of job polarization ... 7

3 Methodology and Data ... 12

3.1 Method ... 12

3.1.1 The International Standard Classification of Occupations (ISCO) ... 12

3.1.2 Measuring job polarization ... 14

3.1.3 Regression model ... 16

3.2 Data ... 17

3.2.1 Occupational data ... 18

3.2.2 Data selection, country classification and sample ... 19

3.2.3 Other data sources and construction of variables ... 21

iv

List of abbreviations

BK Blinder-Krueger index of offshorability

BRIC group acronym for countries Brazil, Russia, India, China GNI Gross National Income

GVC Global value chain

ICT Information and communications technology IPUMS Integrated Public Use Microdata Series IQR Interquartile range

ISCO International Standard Classification of Occupations p.p. Percentage point

pc per capita

PI Price of investment PPP Purchasing power parity

RBTC Routine-biased technological change SBTC Skill-biased technological change SOC Standard Occupational Classification RTI Routine Task Intensity index

1

1 Introduction

Technological developments have consistently contributed to transformations of labor markets throughout the centuries. The industrial revolution and the upcoming of manufacturing industries have affected the nature of work, just as the more recent advancements in Information and Communication technology (ICT) and the ever increasing globalization of goods and service flows. Changes in competitive landscapes, the upcoming of new industries and disappearance of others have led to adjustment processes, such as relocations of workers and shifts in employment structures (Blinder, 2006).

During the past few decades, researchers have observed a previously unknown phenomenon: the polarization of labor markets - an increasing demand for high-and low skill jobs, relative to middle skill occupations, often described as “hollowing out” of the middle (Autor, Katz and Kearney, 2006; Goos and Manning, 2007; Goos, Manning and Salomons, 2009).

A common explanation for labor market polarization is the Routine-biased technological change (RBTC) hypothesis: jobs that consist of repetitive tasks and follow explicit rules can be easily translated into a code and performed by technology. Such routine intensive jobs are often found in the middle of the skill spectrum. Clerical jobs such as record-keeping and calculations, as well as repetitive manufacturing occupations like assembly line-work, typically show a high content of routine-tasks, making them relatively more prone to automation than high- and low-skill jobs that tend to be less routine intensive (Autor, Levy and Murnane, 2003). Evidence that RBTC has contributed to labor market polarization has been found for many advanced economies, such as the US (Autor and Dorn, 2013), 16 Western European countries (Goos, Manning and Salomons, 2014) or Japan (Ikenaga and Kambayashi, 2016).

At the same time the developments in ICT, together with declines in transporting costs have given rise to a trade in tasks. Improved technology allows the coordination of complex production processes at distance, where businesses are able to fragment their production across multiple locations worldwide in Global value-chains (GVC) (Baldwin, 2006). Today firms can easily offshore jobs abroad, to minimize costs, taking advantage of factor cost disparities across countries (Feenstra, 2010; Grossman and Rossi-Hansberg, 2008). Tasks that can be performed at sufficient quality and lower costs abroad, are mostly those that follow explicit rules, do not require no face-to-face contact or abstract thinking (Blinder, 2009). As these jobs are typically also found in the middle of the skill spectrum, offshoring is seen as another contributor to job polarization.

The cause for the emergence of these drivers is seen in the steady decline in costs for technology throughout the past decades (Autor et al., 2003). As this price decline is a global phenomenon, one could assume that RBTC and offshoring are universal forces that affect labor markets around the world in a similar way.

Although a substantial body of literature has studied the phenomenon of job polarization, research so far has largely focused on advanced economies. Less is known about the prevalence of job polarization in less developed countries.

2

I analyze the prevalence of job polarization for a sample of 89 countries and study the effects of drivers that might be related to changes in employment structures.

I conduct my analysis for a much larger country sample than previously done and explicitly differentiate between advanced, emerging and developing countries in the analysis, which – to the best of my knowledge – has not been done before.

Particularly I contribute to the existing literature on job polarization in two manners:

Firstly, I analyze changes in employment structures for the three country groups, which allows identifying patterns and stylized facts on the occurrence of job polarization in relation to economic development.

Secondly, I estimate the effects of typical drivers such as RBTC and job offshoring on occupational changes for the separate country groups by means of a cross sectional regression model. By applying the comparative advantage reasoning for the allocation of production factors of Acemoglu and Autor (2011), I hypothesize and test differing effects across advanced, emerging and developing economies, as I assume the comparative advantage patterns to differ between countries of different development states. I expect labor in less developed countries to hold comparative advantage over machines and offshore labor for a greater amount of tasks than in advanced economies, due to different factor cost structures. Further I estimate the effect of potentially reinforcing or offsetting factors to RBTC and offshoring, namely the effects of Global Value Chain participation, Price of investment and labor market regulation, which I also expect to differ across development states. Currently there is only limited research available that studies job polarization in emerging and developing economies. Maloney and Molina (2016) in a World Bank Working Paper for instance, test the general prevalence of the phenomenon for a small sample of 21 emerging and developing economies, but do not provide a differentiated analysis of drivers in occupational changes for these economies. Furthermore they pool emerging and developing economies together, which does not allow a differentiated analysis of the relationship of drivers of job polarization and economic development.

Reijnders and de Vries (2018) study the effects of trade and technology on the share of non-routine jobs for a sample of 21 advanced and 20 emerging economies and decompose the impacts of these drivers. They find a pervasive effect of technological change on the share of non-routine jobs across advanced and emerging economies, while the effect of task relocation is found to be less strong and works in different directions across countries. However, their analysis is confined to advanced and emerging economies and does not capture the potentially different effects in developing economies. My analysis covers 24 advanced, 19 emerging and 46 developing economies for which I study changes employment shares across 9 occupation classes, based on the 1988 definition of the International Standard of Occupational Classification (ISCO). I use occupational data for the two decades 1990-2010 from the database of the International Labor Organization (ILO, 2016), partly supplemented with data from the Integrated Public Use Microdata Series (IPUMS, 2018).

3

With regards to the drivers of job polarization, I find that RBTC is the predominant effect on occupational changes, both in advanced and emerging economies, with the impact being strongest in advanced economies. The effect of RBTC for emerging economies is robust against the exclusion of agricultural occupations that might be equally driven by structural change. For developing economies I do not find a significant effect of RBTC on changes in employment structures. Further my results do not give evidence for a direction of the impact of offshorability on occupational change for any of the country groups. With regards to potentially reinforcing and offsetting factors to Routine-biased technological change and job relocation, I find that changes in the share of routine occupations are affected by countries’ participation in global value chains, and domestic income levels in developing and emerging countries, while advanced economies appear to polarize regardless of their income level. I find no robust evidence for a relationship of price of investment or labor market regulations and changes in occupational structures.

In a broader context the findings of this paper have implications for policy-makers of countries at all development states. If job polarization turns out to be a continuous phenomenon in both advanced and emerging economies, it could contribute to increasing income inequality with detrimental effects social cohesion and long-term growth. Developing economies might be affected by job polarization in the future. If the price of technology falls sufficiently, RBTC could increasingly affect labor markets in these economies, as this potentially inhibits labor productivity growth and employment opportunities in the manufacturing sector.

The remainder of this thesis is structured as follows: Chapter 2 acts as theoretical foundation of this paper. It introduces the concept of job polarization, and discusses the “task framework” by Acemoglu and Autor (2011) that might explain the mechanisms contributing to the phenomenon. These considerations are used to develop testable hypotheses for the empirical analysis. Chapter 3 discusses the method and data sources used for the analysis. In Chapter 4 I discuss descriptive statistics of important variables that allow inferring stylized facts on occupational changes, as well as the regression results. Chapter 5 concludes and briefly considers some of the economic and political implications of my findings.

2 Theoretical background and hypotheses development

The following chapter introduces the phenomenon of polarizing labor markets, also referred to as job polarization, and reviews literature that examines potential drivers, namely technological change, trade, as well as factors that might reinforce or mitigate their effects. I use both empirical findings and theoretical concepts, to discuss the mechanisms through which these drivers may affect labor market transformations. The considerations from this chapter will be used to develop hypotheses that act as foundation for the empirical analysis of this paper.

2.1 Job polarization

4

and Manning (2003). In an early discussion paper they first document the polarization of the labor market in the United Kingdom, and coined the term “job polarization”. A later version of their work was officially published as Goos and Manning (2007), which will be considered here. The authors document the pattern of employment changes in Britain over the period 1975-1999, by looking at employment shares of different wage deciles. By taking mean wages as simple proxy for occupation groups, they find that there has been a growth in low-paying service occupations, and a significant increase in high-paying jobs, such as professional and managerial occupations, against a decline in middling jobs, such as clerical and skilled manual jobs in manufacturing.

Autor et al. (2006) shortly after that, document a similar development for the United States. They observe that the 1990s were characterized by a more rapid growth of employment in jobs at the bottom and top end of the skill distribution and a relative decline in middle skill jobs, when compared to the 1980s. They show these changes in the employment composition by looking at shifts in the distribution of jobs, based on the mean years of schooling of persons employed in these occupations. They find that changes in employment shares in the 1990s show an inverted U-shape, based on education levels, with employment growth at the extreme ends of the skill level of workers, while the 1980s employment shifts showed a linear relationship with education levels of workers. In a later article, Goos et al. (2014) analyze distributions of labor over time for 16 Western European countries. They track employment shares for 21 occupation groups in the period 1993-2010 and categorize these groups by their mean wage across the countries. Their analysis shows rising shares of high wage jobs of professionals and managers as well as low wage jobs of personal service workers and elementary occupations, while employment in manufacturing and clerical jobs, typically found in the center of the wage distribution, declined at the same time. The paper finds that the phenomenon of job polarization is pervasive across the analyzed European countries.

While the previously mentioned articles are solely focused on advanced economies, a more recent paper by Reijnders and de Vries (2018) analyzes changes in the employment structure for a larger set of 41 countries in the period 1999-2007, comprising both advanced and emerging economies. Using a task based approach and considering both the effects of international trade and technological change, they find that 38 of the 41 analyzed countries were subject to a relative rise in non-routine jobs, while the magnitude of these shifts varies significantly across countries. As routine jobs are typically found in the middle of the skill distribution, while low and high skill occupations are found at the top and bottom, their findings can be interpreted as indication for polarizing labor markets both in advanced and emerging economies.

5

demand effects of labor.1 Although briefly mentioned here, the relationship between job polarization and distributions of income is not the focus of this paper. The contribution to the existing literature on job polarization is the investigation of potential drivers for a larger set of countries that comprises also developing countries, apart from advanced and emerging economies.

2.2 A task framework for the analysis of labor market polarization

A growing body of literature argues that the polarization of labor markets is related to the way in which workplace activities are allocated between capital and labor, and between domestic and foreign workers (Autor, 2013). To provide an explanation for the underlying mechanisms of these allocation processes, Acemoglu and Autor (2011) develop an economic model, which became known as the “task framework”. It describes the assignment of skills to tasks based upon the comparative advantage of different production factors in performing workplace activities. This section introduces this framework, which constitutes the foundation for the subsequent discussion of potential drivers of job polarization and the hypotheses development.

Central to the task framework are the concepts of skills and tasks. The authors define tasks as “unit of work activity that produces output”, and skills as “worker’s endowment of capabilities for performing various tasks” (Acemoglu and Autor, 2011, p. 1045). They argue that it is not skills that produce output, but rather tasks that apply skills. This difference is central for a more nuanced understanding of occupational changes in labor markets, where occupations are understood as bundle of tasks. Workers of a given skill level are able to perform a variety of tasks, where the set of tasks they carry out may change in response to technology and economic costs. The framework uses a Ricardian model of comparative advantage as central explanation for the allocation of tasks. According to this concept, production is assigned to the factor with the lowest economic cost in performing a certain task, which reflects both technological capabilities and opportunity costs (Autor, 2013). It must be noted here that the concept of comparative advantage used in the model is not limited to the cross-country context, but explicitly refers to production factors such as workers of different skill levels and machines, regardless of their geographic location.

The model distinguishes between three types of workers: low-skill (L), medium-skill (M) and high-skill (H) that can perform different tasks, while high-skill levels also represent productivity differences. The production of a final good is composed of a variety of specific tasks y(i), whose output level in turn is determined by the following production function:

𝑦(𝑖) = 𝐴𝐿𝛼𝐿(𝑖)𝑙(𝑖) + 𝐴𝑀𝛼𝑀(𝑖)𝑚(𝑖) + 𝐴𝐻𝛼𝐻(𝑖)ℎ(𝑖) + 𝐴𝐾𝛼𝐾(𝑖)𝑘(𝑖) (1)

where A represents factor-augmenting technologies for each skill level, 𝛼 stands for productivity levels of high, medium and low skill workers in performing a specific task i, and l(i), m(i) and h(i) is the number of workers of the different skill levels performing a task. The last sum of the equation, indicated by the subscript K, stands for capital as separate input factor, which is the autonomous performance of tasks by machines. As the sum signs imply, each task i can be performed by workers

6

of different skill levels, as well as by capital, but the comparative advantages differ across tasks, as captured by different productivity levels 𝛼.

Due to the concept of comparative advantage, in equilibrium, each task will be carried out by the skill group that is most productive, relative to its price.

The tasks performed by a certain skill level are determined by threshold levels, denoted as IL, IM and

IH, where 0 < IL < IH < 1. In equilibrium, all tasks i< IL are performed by low skill workers, all

tasks IL <i< IH by medium skill workers, and all tasks i > IH by high skill workers.

The wage levels of workers are determined by the price level of a certain skill level, and their marginal product, here for the example of low skill workers (as can be done analogously for the other skill levels):

𝑤𝐿 = 𝑃𝐿𝐴𝐿 (2)

The explanatory power of the model lies in the variability of the threshold levels, dependent on technological advances and factor costs. If there is, for instance, a technological improvement for high skill workers - reflected by an increase in AH, the productivity of this skill group will increase, as technology is assumed to be factor augmenting (𝐴𝐿𝛼𝐿). High skill workers will hold a

comparative advantage for a greater range of tasks, resulting in a downward shift of threshold level

IH. As a consequence, high skill workers will perform tasks previously carried out by medium skill

workers, as they lose their comparative advantage to high skill workers. The resulting excess supply of medium skill workers will also reduce IL, due to changes in opportunity costs. In this example, the erosion of medium skilled comparative advantage will reduce medium skilled wages, as this skill group is displaced from tasks they previously performed.

The example - that can be analogously applied to changes in AL or AM - illustrates the effects of technological change on the task allocation across skills and the wage structure within the work force.

Task replacing technologies

A similar mechanism is at work when technology replaces activities that were previously performed by human labor. When the capital productivity 𝛼_𝐾 for a certain task i increases sufficiently, it can be performed at lower cost by machines than by workers of a certain skill group. However, as argued by Autor et al. (2003) not all types of work are equally suitable to machine replacement. In the current era, codifiable routine tasks, primarily performed by medium skilled workers, are most prone to automation, which I will discuss more extensively in the subsequent sections.

7

Offshoring

Alongside task replacing technology, advances in international trade and the international fragmentation of production processes has enabled firms to offshore certain tasks to foreign locations. If labor in an offshore location is sufficiently productive in carrying out a task, it can be attractive to relocate the activity abroad and minimize total costs, where additional transaction costs for the monitoring and coordination of workers must be taken into account.

The task framework by Acemoglu and Autor (2011) is able to formally capture these processes, paralleling the mechanisms described for machines replacing labor.

As I will discuss in more detail in the subsequent section of this chapter, tasks that are most prone to offshoring are typically those performed by medium skilled labor, analogue to those most susceptible to automation. In the model, these workers would perform tasks between the comparative advantage thresholds for high and low skilled labor IL <i< IH. An increasing offshoring of tasks from this category, will displace medium skilled workers from their job and drive IL and IH further apart, where ÎL < IL. Due to a resulting oversupply of medium skilled workers, they will be pushed into tasks that were previously performed by low skill labor.

2.3 Drivers of job polarization

The task framework described above is the foundation for the setup of the empirical analysis of this paper. I will follow the comparative advantage reasoning of Acemoglu and Autor (2011) and combine them with findings from other relevant literature, translating the mechanisms described into testable hypotheses on drivers of job polarization in domestic labor markets.

Routine-biased technological change

When looking at potential causes for the phenomenon of job polarization, Autor et al. (2003) in a prominent article provide a compelling explanation, by taking a closer look at the role of technology in production and the nature of work that constitutes a task.

8

decreased by roughly one-third to one-fourth annually since around 1945, which results in a trillion-fold cumulative decline in computing costs. This price drop in information technology creates an incentive for employers to replace workers with computers. As formally outlined with the task framework in the previous section, the increasing productivity of computers and machines, gives them comparative advantage over human labor in performing routine tasks, if their service is sufficiently cheap.

The routinization hypothesis, also referred to Routine-biased technological change (RBTC), has emerged as leading explanation for the polarization of labor markets.

It constitutes an alternative to a formerly common explanation for changes in the distribution of skills and earnings due to technological change, which is known as the “Skill-biased technological change” (SBTC) hypothesis. This theory only distinguishes between two groups of workers, namely high and low- skill labor, and assumes that technological change leads to a monotone increase in the demand for high skill labor, relative to low-skill labor, as technology is complementary to skills. Although SBTC has been empirically successful for the explanation of earnings distribution in many advanced economies, as shown for the United States (Katz and Murphy, 1992; Autor, Katz and Krueger, 1998), for the UK (Card and Lemieux, 2001) and also for a greater set of OECD countries (Berman, Bound and Machin, 1997), it is not able to explain the “hollowing-out” of labor markets, that was discussed in section 2.1. Specifically, the theory falls short of providing a satisfactory framework for understanding the non-monotone growth of employment by skill level. As shown for the example of the United States for the period 1980 and 2005 by Autor and Dorn (2013) employment changes were strongly U-shaped in skill level, with relative employment declines in the middle of the distribution and relative gains at the tails.

For the phenomenon of polarizing labor markets, RBTC has proven to be an empirically established explanation in advanced economies (for evidence see e.g. Autor and Dorn (2013), Goos et al. (2014), Ikenaga and Kambayashi (2016)). Due to the empirical evidence found for RBTC I hypothesize that due to technological change, the share of routine intensive jobs declines in advanced economies:

Hypothesis 1a: In advanced economies, the employment share of routine-jobs declines, relative to non-routine jobs.

However, I expect the replacement of routine-intensive jobs with capital to be more attractive for firms in advanced economies, where the costs of labor are higher, while in emerging and developing countries it might be still more attractive to have routine tasks executed by human labor. Prior research by Reijnders and de Vries (2018) finds evidence for RBTC in emerging economies, but I expect a less pronounced effect, compared to advanced economies as human labor is expected to hold comparative advantage for more jobs, due to lower labor costs and consequently a higher price of capital, relative to labor, which should, according to the task framework decelerate the replacement of tasks by technology.

9

For developing economies I expect the effect of RBTC to be insignificant, as I assume labor costs relative to capital to be even lower than in emerging economies. Following the reasoning of the task framework, this would prevent an erosion of the comparative advantage of human labor for the performance of routine tasks. Furthermore, the employment structure in developing economies typically consists of a lower share of middle skill workers performing routine tasks that would be susceptible to automation. The workforce in many of these countries often has a high share of low skill service and artisanal production jobs, where workers perform abstract and non-routine manual tasks that cannot be easily automatized (Maloney and Molina, 2016).

Hypothesis 1c: There is no significant decline in the share of routine-jobs, relative to non-routine jobs in developing economies.

International trade and offshoring of tasks

Similar to the automation of tasks, there are comparable effects at work for the offshoring of labor. Firms have an incentive to relocate tasks abroad, when they can be carried out at lower costs in other countries, as mentioned in the discussion of the task framework above. Declines in transport and communication costs have facilitated the spatial separation of production processes, and the development of global value chains (GVC) (Baldwin, 2006). Trade is no longer a simple exchange of final goods, but increasingly involves the execution of production processes in locations where they can be conducted most efficiently. Through advancements in information technology, firms can deliver instructions instantaneously and ship components and intermediate products at low cost, which allows producers to take advantage of factor cost disparities between countries. The consequence has been an increasing offshoring of both manufacturing and service activities, which Grossman and Rossi-Hansberg (2008) refer to as “trade in tasks”. However, not all types of occupations are equally prone to offshoring. Some jobs require the physical presence of workers, such as performing surgery or driving a taxi. Other activities such as customer service in call centers or production steps in manufacturing can be relocated abroad. Blinder and Krueger (2013) call the susceptibility of jobs to be moved abroad “offshorability”.

10

countries can be expected to be rather the destination of relocated tasks, as they are more likely to hold comparative advantage in performing offshorable tasks. According to this logic, we could expect a decrease of routine jobs in advanced economies, which could drive a polarization of jobs, while their number increases in emerging and developing economies, which would offset polarization forces. However, not all emerging and developing economies are equally the receivers of relocated jobs. As Reijnders and de Vries (2018) show, the relationship of task relocation on occupational shares of routine jobs is quite diverse across emerging economies, therefore I expect of offshoring to be invisible at an aggregate level in both developing economies and emerging economies.

Hypothesis 2a: In advanced economies, a higher offshorability of an occupation leads to a relative decline in its employment share, relative to occupations with a lower offshorability.

Hypothesis 2b: The offshorability of occupations does not have a significant effect on employment shares of occupations in developing and emerging economies.

It should be noted here, that the routine intensity and offshorability of jobs merely describes their propensity of being affected by automation and task relocation, based on occupation characteristics. Whether jobs are actually automated or offshored may depend on additional factors that vary at the country level. Although the distinction between different countries of different development stages already allows a certain differentiation of effects for different country groups, I will discuss three potential drivers that might be related to changes in routine-occupations. These could be considered as reinforcing and mitigating factors to the two drivers discusses above.

GVC participation

As reasoned in the discussion about the offshoring of jobs, not all developing and emerging markets can be expected to equally be the recipient of jobs that are relocated abroad, just because labor costs are low. Firms will likely base the location decision for the destination for the offshoring of jobs on a variety of determinants, which besides costs may include factors such as local skill levels, the quality of infrastructure, risk profiles or the general business environment (Oshri et al., 2015). In a recent article, Timmer, Miroudot and de Vries (2019) show that facilitated by trade in tasks, countries increasingly specialize in certain functions of value chains. They find a large heterogeneity in specialization patterns across countries of similar income levels. Fabrication activities for instance, are concentrated in China and a limited number of other emerging economies, while India specializes in service activities. Advanced economies on the other hand rather specialize in activities like management or marketing.

11

employment of non-routine jobs, as GVC participation reveals to what extent countries vertically specialize according to their comparative advantage. Whether routine jobs increase, because countries are origin or destination of offshored jobs, will again depend on their development state.

Hypothesis 3a: A higher GVC participation leads to an increase in non-routine jobs in advanced economies.

Hypothesis 3b: A higher GVC participation leads to a decrease in non-routine jobs in developing and emerging economies.

Price of investment

As outlined in the task framework, the replacement of labor through machines and computers depends to a great extent on the factor price of capital, relative to labor. Capital will only have comparative advantage in the performance of tasks, if it is sufficiently cheap. Karabarbounis and Neiman (2014) show that the share of labor income in many countries declined significantly since the early 1980s, which they attribute to lower prices for capital through technological advances. They find that a relative decline in the price of investment goods is able to explain roughly half of the decline in labor income shares, as capital becomes cheaper for firms. However, Dao et al. (2017) show that the decline in the relative price of investment has been predominantly an advanced economy development, where it declined by about 12 percent between 1993 and 2014. In emerging and development it declined at a lower overall rate of 7 percent in the same period, although with a significant dispersion of price changes across countries: in some emerging and developing economies prices fell, in others they stagnated or even increased. As the price of investment is a central determinant of the substitution of labor for capital, its price is likely to affect firms’ incentives to automatize jobs. I assume that routine jobs are more likely to be replaced by capital, where the price of investment for capital is low.

Hypothesis 4: A lower price of investment is related to an increase in the share of non-routine jobs.

Labor market regulation

Another factor that might influence the magnitude of job polarization in labor markets is the institutional environment of the labor market, which may influence the freedom at which firms can use machines and workers in different combinations or relocate jobs. The flexibility in the dismissal of workers is substantially determined by the labor market regulations to which employers must adhere. With the relaxation of regulations, enterprises have greater freedom in hiring and firing, to increase or decrease wages and in the assignment of tasks and replacement of workers whose skills have become obsolete (Piore, 1986; Regini, 2000). Bound and Holzer (2000) find that labor mobility tends to be lowest for non-college workers, which are precisely those affected most by automation and offshoring. If these workers are easily laid-off by their employees, and cannot find a new job, suitable to their qualification, the polarization of the labor market might be more pronounced.

12

3 Methodology and Data

This chapter describes the methods applied in the analysis of this paper and discusses the data sources. Based on the hypotheses established in the previous chapter, I empirically test the relationship between potential drivers and changes in occupational structures by means of two regression models. The analysis is conducted for a sample of 89 countries that consist of 24 advanced, 19 emerging and 46 developing economies.

3.1 Method

I study job polarization patterns across countries and its potential drivers in two steps: First I look at employment shifts between different occupational categories in the two decades 1990-2010 and compare the patterns of advanced, emerging and developing economies to identify broad trends that allow inferring stylized facts about the polarization propensity of countries, based on their development state. In a second step I apply two different cross sectional regression models to test the relationship between potential drivers of job polarization and occupational shifts. I begin the section with an introduction of the International Standard Classification of Occupations, which is the categorization system of occupations used in this paper and builds the foundation of the empirical analysis. This is followed by a discussion on the measurement of job polarization and the regression models I use to test the hypotheses established in the previous chapter.

3.1.1 The International Standard Classification of Occupations (ISCO)

The polarization of labor markets can be analyzed by looking at changes in the employment composition of occupations. As described in section 2, occupations can be seen as bundles of tasks, which vary in their susceptibility to automation and offshoring. When it comes to the classification of occupations, the International Standard Classification of Occupations (ISCO) has emerged as a common standard that has been implemented by many countries around the world. Developed by the International Labor Organization (ILO), it facilitates the international comparison of occupational statistics, through the establishment of a standard framework that allows categorizing jobs into defined sets of groups. The classification of occupations is based on the nature of work in the tasks and duties involved. To account for changes in the nature of work, the system has been revised and updated over time. The first two generations ISCO-58 and ISCO-68 were introduced in 1958 and 1968, respectively, being superseded by the third generation, ISCO-88 in 1988. ISCO-88 constituted a significant break from the first two generations in methodology, as it first used the concepts of skill level and skill specialization as criteria to arrange similar occupations into groups, while the previous classifications had a stronger focus on the goods and services produced (Elias, 1997; ILO, 2012). Based on the type of work performed, ISCO-88 first groups jobs whose main tasks show a high degree of similarity into one occupation category. The second classification criterion “skill”, describes the skill level, which is the complexity and range of tasks involved, as well as the skill specialization, which is defined as the field of knowledge required, the tools, machinery and materials used, as well as the goods and services produced (ILO, 2004a).

13 Table 1: ISCO-88 skill levels and required education

Skill level Corresponding education/qualification

4 Tertiary education (begun at ages 17-18, lasting 3-6 years and leading to university degree or equivalent) 2 Tertiary education (begun at ages 17-18, lasting 3-4 years, but not giving equivalent of university degree) 3 Secondary education (begun at ages 11-12 and lasting 5-7 years)

1 Primary education (begun at ages 5-7 and lasting approximately 5 years)

Source: Elias (1997)

These four skill levels are used to define the broad classification structure of occupations, which leads to ten major occupation groups, numbered by single digits. Each major group can be further divided into a total of 28 two-digit sub-major groups, 116 three-digit minor groups and 390 four-digit unit groups. The ten major occupations and their sub-divisions are depicted in Table 2.

Table 2: ISCO-88 major groups and subdivisions with individual RTI and BK scores

Major

Group ISCO code Subdivision

ISCO Skill level

RTI BK

1 Legislators, senior officials and managers

11 12 13

Legislators and senior officials Corporate managers General managers 4 -0.57 -0.65 -1.45 -0.42 -0.13 -0.46 2 Professionals 21 22 23 24

Physical, mathematical and engineering science professionals

Life science and health professionals Teaching professionals Other professionals 4 -0.73 -0.91 -1.47 -0.64 1.31 -0.59 -0.70 0.43 3 Technicians and associate professionals 31 32 33 34

Physical and engineering science associate professionals Life science and health associate professionals

Teaching associate professionals Other associate professionals

3 -0.29 -0.23 -1.37 -0.34 0.08 -0.58 -0.65 0.31 4 Clerks 41 42 Office clerks

Customer services clerks 2

2.41 1.56

0.62 -0.06 5 Service, shop and _{market sales workers} 51 ₅₂ Personal and protective services workers _{Models, salespersons and demonstrators} 2 -0.50 _0.17 -0.78 _-0.73 6 Skilled agricultural and _{fishery workers} 61 ₆₂ Market-oriented skilled agricultural and fishery workers _{Subsistence agricultural and fishery workers} 2 0.14 _0.47 -0.84 _-0.84

7 Craft and related trades workers

71 72 73 74

Extraction and building trades workers Metal, machinery and related trades workers

Precision, handicraft, printing and related trades workers Other craft and related trades workers

2 -0.08 0.58 1.74 1.38 -0.77 -0.27 1.94 1.41 8

Plant and machine operators, and assemblers

81 82 83

Stationary-plant and related operators Machine operators and assemblers Drivers and mobile-plant operators

2 0.45 0.62 -1.42 1.87 2.66 -0.84 9 Elementary occupations 91 92 93

Sales and services elementary occupations Agricultural, fishery and related laborers

Laborers in mining, construction, manufacturing and transport 1 0.14 0.38 0.57 -0.64 -0.84 -0.48 0 Armed forces - - - -

Notes: Occupation overview from Elias and Birch (1994), RTI and BK values from Goos et al. (2014).

14

ISCO-88 has been retained in ISCO-08, ILO does not recommend to link the two standards for empirical analyses, due to significant changes in the sub-categories (ILO, 2012). As countries are only slowly adapting to the ISCO-08 generation, data availability is still highly limited for this standard. Furthermore, the great majority of recent literature, which serves as orientation for this paper, works with ISCO-88. Measures for the characteristics of the occupational groups, which constitute an integral part of my analysis, are therefore only available at the ISCO-88 level. For these reasons ISCO-88 is the preferred classification standard for my analysis.

3.1.2 Measuring job polarization

By using an occupation classification system such as ISCO-88, it is possible to identify patterns of polarization in labor markets, which can be done in different ways. Pioneering research by Autor et al. (2003) distinguishes between four categories of tasks that are considered inputs to occupations, namely non-routine analytic, non-routine interactive, routine cognitive and routine manual. The authors then track changes in the shares of task inputs across industries for the period 1970-1998 in the US. They find a consistent increase in the input share of routine cognitive and routine manual tasks. Reijnders and de Vries (2018) follow the categorization of tasks by Autor et al. (2003) and group occupations into the same four categories, based on their task intensity. They show total changes in the shares of non-routine occupations over time for different countries, where the documented increase in the share of non-routine jobs for most countries can be interpreted as indication for labor market polarization. Goos et al. (2014) choose yet another approach, showing changes in the employment share of two-digit occupations groups over a period of 18 years, but order them by mean European wage to show the polarization of labor markets. For the present paper the heterogeneity of the country sample does not allow a ranking of jobs according to mean wages, neither does the low data availability for many countries allow a calculation of total changes of non-routine jobs over an extended period of time. I calculate annual changes in the employment share of occupational groups, to analyze patterns for advanced and emerging and developing economies, visualized by boxplot graphs in chapter 4 of this paper.

Measures Routine intensity and offshorability

In a first step I use the task categorization from Autor et al. (2003), and group two-digit ISCO-88 occupations accordingly. The grouping is depicted in Table 3. For the analysis of potential drivers of job polarization, it is necessary to categorize occupation groups based on their routine intensity.

Table 3: Classification of ISCO-88 occupations

Routine Non-routine

analytical/ interactive

Clerks (41-42)

Personal and protective services workers (51)

Legislators, senior officials and managers (11-13) Professionals (21-24)

Models, salespersons and demonstrators (52)

manual

Skilled agricultural and fishery workers (61-62) Extraction and building trades workers (71) Plant and machine operators, and assemblers (81-82)

Technicians and associate professionals (31-34) Elementary occupations (91-93)

Craft and related trades workers (72-74) Drivers and mobile-plant operators (83)

15

As occupational data for most countries is only available at the one-digit level, the categorization must be converted from the two-digit to the one digit level, which I will further discuss later in this section.

The routine intensity of occupations is most commonly measured by the Routine Task Intensity index (RTI) as constructed by Autor and Dorn (2013). The index uses the composition of tasks that make one occupation, by measuring the content of routine (R), manual (M) and abstract (A) tasks involved (see equation 3).

𝑅𝑇𝐼_𝑘 = ln(𝑇_𝑘𝑅_{) − ln(𝑇}

𝑘𝑀) − ln(𝑇𝑘𝐴) (3)

A positive RTI value indicates a routine intensive occupation, a negative value a non-routine occupation. The highest routine intensity is that of office clerks (41) with a routine score of 2.41, the lowest is that of managers of small enterprises (13) with a score of -1.45. It should be noted here that the values are relative measures, where concrete values may differ between data sources, based on the standardization methods.

For the offshorability of tasks I use the findings of Blinder and Krueger (2013), who assess the offshorability of jobs by three different methods: Firstly by asking workers directly on the difficulty of performing their job from abroad, secondly by inquiring respondents about the nature of their work and thirdly by professional coders’ assessment of the offshorability of jobs. They find that the latter method provides the empirically most accurate assessment of offshorability, which is why I follow their assessment and make use of that measure. For the Blinder and Krueger measure of offshorability I will refer to as BK in the following. Concrete values of both RTI and BK for the ISCO-88 classification are available from the Online-Appendix of Goos et al. (2014).2 Detailed measures are included in Table 2 of section 3.1.1.

As the available measures for routine intensity and offshorability are available at the two digit level, I map RTI and BK values to the one-digit ISCO-88 level. In the case of most major groups this would be readily done, as there is not much heterogeneity within major groups, which would allow taking an unweighted average value of the values for each occupation, without risk of significant bias for an empirical analysis. Such homogenous groups are Major group 1-4, 6 and 9. In the case of group 7 and 8, however, sub-major groups show significant variations in their routine intensity. Occupation 71 Extraction and building trades workers are routine, while the rest of the craft and related trades

workers (72-74) are non-routine. The same applies to major group 8, where occupation 81 and 82 of plant and machine operators, and assemblers are routine, while drivers and mobile-plant operators

are non-routine (see Table 2 for an overview). Taking unweighted average of these RTI values would lead to inaccurate numbers, as the sub major occupation groups are likely to differ in their employment shares. As detailed occupational data at the two-digit level is only scarcely available, I use occupational data for an example country to correct for this potential bias. I take occupational shares at the two-digit level from Germany in 2010 to compute weighted averages of the RTI and offshorability scores in each major group, based on the employment distributions within each one digit occupations. The weighted averages of the RTI and BK values in each major group are depicted

2

16

in Table 4. To control for unique characteristics of the German labor market, I do a second calculation with employment shares of the UK, with openly available occupational data from the Office for National Statistics (ONS, 2019). The occupational data is converted from the Standard Occupational Classification 2000 (SOC2000) to the ISCO-88 system by means crosswalk data from Lambert and Prandy (2018). Figure A.2 in the appendix depicts the correlation between the two calculations, showing that they show now significant deviation. I thus use the values obtained in the calculation with the data from Germany. Table 4 below depicts the obtained RTI and BK values for each ISCO-88 major group. The correlation coefficient between the composite values for RTI and BK at the one digit level is 0.17 and statistically significant. The value is much lower than that at the two digit level individual, which Goos et al. (2014) report to be 0.46 and significant. The interrelatedness of routines and offshorability is therefore not entirely captured in the composite measures, which is likely to affect the accuracy of my analysis.

Table 4: Routine intensity and offshorability of ISCO-88 one-digit occupations

Major Group RTI BK

1 Legislators, senior officials and _managers -0.94 -0.28 2 Professionals -0.86 0.37 3 Technicians and associate _{professionals} -0.40 0.03

4 Clerks 2.33 0.56

5 Service workers and shop and _{market sales workers} -0.27 -0.76 6 Skilled agricultural and fishery _workers 0.14 -0.84 7 Craft and related trades workers 0.50 -0.13 8 Plant and machine operators, _{and assemblers} -0.45 0.76 9 Elementary occupations 0.32 -0.59

Notes: Numbers are weighted averages of RTI and BK measures, constructed with data from Goos et al. (2014)

3.1.3 Regression model

To empirically investigate the drivers of polarizing labor markets, I apply two cross-sectional regression models. In my analysis I generally follow the approach of Goos et al. (2014), who use the routine intensity and offshorability of tasks as explanatory variables for changes in employment shares per occupation. In an economic model they formally describe the impact of jobs’ routine intensity and offshorability as affecting the factor costs for producing one unit of task i which can be modelled with the following equation:

𝜕𝑟𝑖

𝜕_𝑖 = 𝛾𝑅𝑅𝑖 + 𝛾𝐹𝐹𝑖 (4)

where r captures the costs for inputs for the production of a unit of task i, other than domestic labor, and 𝑅_𝑖 and 𝐹_𝑖 are the routine intensity and offshorability of task i. As both R and F are increasing with the level of routineness and offshorability, the expectation is that 𝛾𝑅 < 0 𝑎𝑛𝑑 𝛾𝐹 < 0. As a

17

I empirically test this relationship with a model that estimates changes in occupational shares across the 9 ISCO-88, which is specified as follows:

∆𝑂𝑆_𝑗𝑐 = 𝛽₁+ 𝛽₂𝑅𝑇𝐼_𝑗+ 𝛽₃𝐵𝐾_𝑗+ 𝜀_𝑗𝑐 (5) where the dependent variable ∆𝑂𝑆_𝑐𝑗 is the average annual percentage point (p.p.) change in the employment share of occupation j in country c during the two decades 1990-2010, RTI is the routine task intensity, BK is the Blinder and Krueger measure of offshorability. The model is estimated first for a group of Western European countries that follows the classification of Goos et al. (2014) to allow the comparison of results and verify the model setup, and then separately for advanced, emerging and developing economies, to test for differing effects of the explanatory variables for the country groups.

The variables of the first model (equation 5) are occupation specific. As the additional potential drivers of occupational changes discussed in hypothesis 3-5 are country specific, they cannot be estimated in the same model. For this reason I set up a second model, where I estimate the effect of factors that might affect the employment share of non-routine jobs, which I specify as follows:

∆𝑆𝑁𝑅_𝑐 = 𝛽₁+ 𝛽₂𝐺𝑉𝐶_𝑐 × devstate+𝛽₃𝑃𝐼_𝑐 × devstate+𝛽₄𝑙𝑎𝑏𝑜𝑟𝑐𝑜𝑠𝑡𝑠_𝑐 × 𝑑𝑒𝑣𝑠𝑡𝑎𝑡𝑒 +

𝛽₄𝐿𝑀𝐹_𝑐 + 𝜀_𝑐 (6)

where the dependent variable ∆SNR is the p.p. change in the share of non-routine jobs, GVC is the participation in Global Value Chains, PI is the relative price of investment goods, laborcosts is a control variable measuring per capita income levels and wages and LMF a measure of labor market flexibility. For the first three explanatory variables, I include interaction terms with the development state of countries that takes a level of 1 for advanced economies, 2 for emerging economies and 3 for developing economies.

I choose a cross-sectional model for my analysis, as it gives the highest flexibility in the use of the available data. It allows including a much larger set of countries that partly shows incomplete time series in the period 1990-2010, particularly in the case of emerging economies. The estimation of a panel data model would requires complete time series for each country, which would require the imputation of values for missing periods. In the case of this paper, the imputation of occupational shares for developing countries would likely lead to an overestimation of changes over time. Another argument for the use of a cross sectional model is the nature of most of the selected explanatory variables. Routine intensity and offshorability are constant by definition and most other variables can be regarded as relatively stable, which allows a meaningful cross country comparison without accounting for changes over time.

3.2 Data

18

3.2.1 Occupational data

Two commonly used databases offer harmonized international occupational data according to ISCO-88: The Integrated Public Use Microdata Series (IPUMS) by the Minnesota Population, and ILOSTAT, the database of the International Labor Organization. IPUMS consists of harmonized census micro-data from around the world, having collected a large archive of census samples that contains occupational data for a total of 93 countries. However, the database relies solely on population censuses, conducted at different points in time that show significant discrepancies in time coverage between countries. For this reason the database of the International Labor Organization (ILO) is chosen as main data source of occupation data for this paper. It offers an extensive dataset on annual employment, containing statistics from national sources on employment by occupation, containing both nationally reported, harmonized estimates, which offers comparable information across countries and time.3 Although national databases also contain information on occupational structures, the data collection methods and classification systems are not always in line with each other, making a cross-country comparison difficult. Due to the scope of this paper, a harmonized data source that has been constructed based with a unified methodology and occupational classification is the preferred source of information.

In the ILOSTAT database, employment considers all persons of working age, which is mostly defined as persons aged 15 or older, unless national laws and practices vary from this threshold level. Employment is reported for paid employment and self-employment, whether at work or not. The database contains information from labor force surveys, population censuses, and household surveys, where the source of information is always indicated in the data. ILO (2018), in its methodology overview, recommends the use of labor force surveys. These yield representative data for the entire population of a country, covering all branches of economic activity, economic sectors and all categories of workers, including self-employed, family workers, casual workers, and multiple jobholders, where official statistics in the latter case typically only consider the primary job. The information obtained from household surveys and population censuses on the other hand is described as less reliable for the purpose of occupational statistics, as they typically do not allow enquiring details on labor market activities of the respondents. Other sources such as establishment surveys or administrative records in some cases may provide information on employment by occupation, but they do no not cover the entire employed population, as they typically exclude the informal economy, small establishments and some specific economic activities or occupations (ILO, 2004b, 2016). In many cases the ILOSTAT database contains several types of sources for the same country. I follow the ILO recommendations and prioritize labor force surveys for my analysis, which constitutes the majority of data. To include as many countries as possible, I make use of population census data if labor force surveys are not available for a country, followed by household surveys. To avoid bias in the data, I always use only one type of source per country. ILOSTAT reports occupational data both for persons employed and as distribution across occupations. The classifications are available for the ISCO generations 68, 88 and 08, according to the implemented standard at the respective time. Since 2008 countries have progressively adapted their data reporting to the most recent ISCO-08 standard. However, the ISCO-88 generation is still the most commonly

19

used classification, yielding a higher availability of data than ISCO-08. ILOSTAT comprises data for 152 countries at ISCO-88, and 142 countries at ISCO-08, however, the data coverage across countries and years is much higher for ISCO-88, which makes it the preferred data unit for this paper, as already discussed in the previous section.

3.2.2 Data selection, country classification and sample

ILOSTAT captures occupational structures at the one-digit level. For the unit of measurement there is a trade-off between the granularity of occupational descriptions and the breadth and quality of data: The more detailed the description of occupations, and thus, the more nuanced the ISCO groups, the more accurate are the predictions on the susceptibility of jobs to automation and offshoring. Very detailed categorization into ISCO minor-groups, however, will greatly limit the number of countries which offer such information. Furthermore the accuracy of these numbers across countries is questionable, when task descriptions become too detailed. For the objective of this paper, one digit occupations give sufficient detail of occupational characteristics, allowing a meaningful analysis of occupational changes. At the same time they allow the integration of a large sample of countries. Other researchers such as Das and Hilgenstock (2018) or Maloney and Molina (2016) have used one-digit levels of occupations for studying polarization patterns for a greater sample of countries. Maloney and Molina (2016) show for the example of the US that the polarization patterns at using 1-digit occupations are consistent with the polarization results obtained by Autor (2010), who uses more detailed occupational categories.

20

Finally the ILOSTAT dataset contains some irregularities in the times series of some countries, which is indicated as ‘break in series’ and mainly due to revised methodologies in surveys. For instance the times series of Guatemala, that contains only observations for the years 2003 and 2004, reports a decline of 6 percent for occupational group 6 within one year and an increase of occupation group 9 by 5.9 percent at the same time. North Macedonia, in one survey reports a decline of 11.8 percent of occupation 6 in one survey and Bangladesh an annual decline 4.8 percent in the employment share of group 5 within 3 years. As also indicated in the dataset, these extreme values are driven by changes in the classification methodology, rather than actual employment shifts. I find that these extreme values significantly affect the distribution of my dependent variable “annual p.p. in occupational employment share”. Therefore I correct for outliers in the dependent variable, by removing observations that deviate more than two standard deviations from its mean. Further I completely drop countries that contain outliers for three or more occupations. As the employment shares of occupations are interconnected, I assume the other reported shares to be equally inaccurate, if three or more show extreme values. As Figure A.1 in the appendix shows, the outlier correction brings the distribution of the variable closer to normality, making in suitable for an OLS estimation. Although there is still a certain kurtosis observable, the correction should increase the validity of the estimations.

ILOSTAT does not contain data for China or the United States at the ISCO-88 level. To include these two countries in my analysis, I supplement my data set with information from IPUMS (2018). For the analysis I categorize countries into three categories: Advanced, Emerging and Developing. In order to relate my results to existing research, I follow the county classification of Reijnders and de Vries (2018) for advanced and emerging economies, and classify the remaining countries generally as “developing”. A classification of countries will always require arbitrary choices. Some countries that I classify as developing, as for instance Thailand and Malaysia, could be just as well regarded as emerging economies, but that would take away from the comparability of my findings. Furthermore, single countries affect the results of my analysis with equal weights, which is why I expect the impact of single economies that could be categorized differently to be limited.

The sample of countries that Reijnders and de Vries (2018) categorize as advanced and emerging economies is based on the country set included in World-Input-Output Database (WIOD), which together account for approximately 85 percent of World GDP (Dietzenbacher et al., 2013). By incorporating a large set of additional countries in the country sample, my analysis covers the vast majority of the world economy.

Alternative classifications, such as the IMF categorization into advanced, emerging and low-income countries, would reduce the number of low-income countries to a number of 9 countries for which occupational data is available from ILOSTAT. This would significantly limit the explanatory power of my analysis. The World Bank categorization into high- , upper middle -, lower middle- and low income countries is merely based on per capita income and does not take the economic structure or broader development into account. I therefore conclude that the classification of Reijnders and de Vries (2018) is preferred for the purpose of this paper.

21

3.2.3 Other data sources and construction of variables

In this section I discuss the data sources and construction of the variables used in the regression models.

Dependent variables

The dependent variable of the first regression model measures the mean percentage point change in the employment share of an occupation in a country per year, in the two decades 1990-2010. This gives one value per country, that is used for the cross sectional regression. As I assume relatively stable developments in employment structures, with little volatility between years using calculating a single value for each country should yield meaningful results for the purpose of this paper.

The dependent variable of the second regression model measures the aggregate annual employment share in non-routine occupations, by adding together the changes of non-routine occupations (negative RTI value) 1,2,3 and 8 together.

Routine intensity and Offshorability

The measures for the routine intensity of tasks (RTI) and their Offshorability (BK) are taken from Goos et al. (2014) and mapped into one digit occupation groups, as discussed before. Values are normalized to have zero mean and unit standard deviation. The data also contains a measure for occupational group 6: skilled agricultural and fishery workers, characterized by a positive RTI value. Most studies on job polarization exclude this occupational group from their analysis, as they do not constitute a significant employment share in advanced economies. When studying developing countries, however, they should be part of the analysis, considering the typically higher employment share of this group Herrendorf et al., (2014).

Global Value Chain participation

For the effect of trade on job polarization, I test for the impact of countries’ participation in Global Value Chains, measured by their vertical specialization with data from Pahl and Timmer (2019). The dataset gives information on the domestic value-added content in the exports of manufacturing goods (VAX-D ratio) for 91 countries over the period 1970 to 2013. The ratio ranges from 0 to 1 where a low value indicates a high GVC participation and a high value a low participation level. For a more intuitive interpretation of the measure I transform the values to 1-VAXD, which results in a high value for high GVC participation and vice versa. I take an average of the values over the two decades 1990 and 2010 Global Value Chains to have an indication of countries’ GVC participation.

Price of Investment

22

fabricated metal products and motor vehicles (World Bank, 2015, pp. 251–260). Although the PPP correction limits the direct comparability of the values between countries, it gives an indication of the overall price level in a country, which I expect to be related to the costs of labor, which makes it a useful indicator for this paper. The data is given as a ratio, relative to the US (US=1) for 168 countries and is only openly available for 2011. Although this is outside of the period under analysis of this paper, it should still give a useful indication for a country comparison.

Labor market flexibility

To measure the flexibility of the labor market, I use the Labor Market Regulation index, which is part of Economic Freedom Dataset by the Fraser Institute (2018), and is composed of the five inputs

Hiring regulations and minimum wage, Hiring and firing regulations, Centralized collective bargaining, Hours Regulations, Mandated cost of worker dismissal, Conscription and Labor market regulations. An overall value of 10 is the highest possible index score, a value of 1 the lowest. In

order to earn high marks in the rating, countries must allow market forces to determine wages and establish the conditions of hiring and firing. The index is consistently reported from 2000 onwards, covering an extensive sample of 162 countries. I use the average index score between 2000 and 2010 as measure for labor market flexibility.

Control variables

I use per capita income and domestic wages as control variables for labor costs in the second model. For per capita income I take Gross National Income data from the World Bank, based on the Atlas method. The Atlas method uses a conversion factor to reduce the impact of exchange rate fluctuations in the cross-country comparison of national incomes by taking a country’s exchange rate for that year and for the two preceding years, adjusted for the difference between the rate of inflation in the country and international inflation. I take data at per capita level in current USD from World Bank (2019). As annual per capita levels for my country sample range between USD 160 and USD 63,790 I normalize values by means of a log-transformation.

For wage levels I use data from the World Development Report (2013), which covers the period from 1983 – 2008, for a number of 171 countries. The dataset reports wages for 162 occupations and various industries. The classification of occupations, however, follows an individual system used in the report, which does not allow mapping these numbers to the ISCO-88 standard. To obtain values suitable a for cross country comparison, I use hourly wages in uniform calibration in US-dollar and take the median wage across occupation in 2005, which gives an indication of the relative wage level of a country. Due to the high variability in values, I use the natural logarithms of the variable.

4 Results