Changes in Employment based on the Education Criterion in the EU: Skill Upgrading or Skill- Polarization?

This paper focuses on changes in the structure of the labor market, the main criterion being education, which determines the skill level of each worker. This allows considering highly educated workers as high-skilled workers. Due to the availability of the data set from EU KLEMS, we focus on all EU countries between 2008 and 2015. This contribution resolves the question, what the trends within jobs are in different industries of EU countries based on the education criterion. The results imply (with exceptions) skill-upgrading across all industries and in most countries (only 5 countries stagnate). Proportions of high-skilled workers increased in the countries concerned by about 3–10 p. p. Similarly, wage growth was proportional to skills. All this provides evidence about the skill-upgrading process (which is consistent with the SBTC hypothesis), rather than skill-polarization. Some evidence on skillupgrading raises the question how the growing number of high-educated workers have affected productivity. The results show that in most cases there is a direct relationship, but there are country and industry exceptions.


Introduction
The constantly changing demand structure does not allow for a reliable prediction of future changes in the employment structure. Additionally, the pace of technological progress and related changes cannot be predicted. Occupations that were twenty or thirty years ago a privilege of high-skilled workers, and which usually required a higher level of education (e.g. accountants), can now be replaced with software or even workers without a higher level of education.
Changes in the employment structure are usually attributed to the role of technological change, globalization and institutions (Goos et al., 2009). In the 1980s and 1990s, it was believed that new technologies had the greatest impact on the structure of employment 1 , resulting in the formulation of the skill-biased technological change (SBTC) hypothesis (Autor & Katz, 1999). This hypothesis, described in the next section, assumes a direct relationship between the demand for workers and the level of their skills; high-skilled workers are demanded more, while workers with lower skills are demanded less. The result is skillupgrading. Against the theoretical assumptions, however, is a picture of reality that does not fully conform to these assumptions.
In the last fifteen years, contributions 2 have appeared that described a change in developmental tendencies in labor markets that does not correspond to the SBTC hypothesis. This change is job polarization 3 . The term "job polarization" comes from Goos & Manning (2007), who thus called a situation where the numbers of high-skilled workers and low-skilled workers are growing, while the number of middle-skilled workers is declining; the result of this situation is skill-polarization.
This paper focuses on a change in the structure of the labor market where the main criterion is education, which determines the skill level of each worker (Acemoglu 1998, Krueger 1993, Michaels et al., 2014. This fact allows considering high-educated workers as high-skilled workers etc. In this paper we use the interchangeability of these terms. This also allows us to contribute to the discussion on the current labor market issue: the collision of theoretical assumptions (skill-biased technological change hypothesis) with reality (job polarization). The research question is: What are the trends of employment structure in individual sectors of EU countries based on the educational criterion? We are exploring whether there is skillupgrading or skill-polarization. The results are presented both at sectoral level and the country level. Given the availability of the data set, the examined period is between 2008 and 2015. Changes in employment structure in terms of education are compared with changes in productivity, expressed as value added per employee, and with changes in wages.
The results of the contribution show that the evolution of the labor market structure in EU countries confirms the SBTC hypothesis. The increased proportion of high-skilled workers, found by this contribution as regards highly educated workers, is a trend across almost all industries, while their reduced proportion is rather exceptional. This supports the SBTC hypothesis, which assumes an increasing number of high-skilled workers. Similarly, wage growth was proportional to skills. Some evidence about the skill-upgrading process raises the question: How did the changing structure affect productivity? The results show that the growing number of educated workers (mostly high-educated) has affected productivity positively. The expected direct relationship is common in most cases (in most countries and most industries), but there are country and industry exceptions.
The rest of the contribution is structured as follows. Section 2 describes the current structural changes that are taking place on the labor market, in particular SBTC hypothesis and polarization on the labor market. Section 3 looks at the specification of data sources and measurement issue. Section 4 shows how the structure of employment has changed, when regarding the level of education as the main criterion. The question how it affected the productivity and wages is also addressed. Conclusion is presented in section 5. Sometimes we can see the term "missing middle" (Oldenski, 2014), which basically reflects the negative effect of job polarization. Card & DiNardo (2002, p. 734) described SBTC as follows: "burst of new technology caused a rise in the demand for highly skilled workers". High-skilled workers generally need more new technologies to work, implying not only a complementary relationship between human capital and new technologies but also a direct proportion: involving more and more new technologies requires more high-skilled workers to work with them (or who would only operate the new technology). A reverse relationship also applies: new technology usually requires more skilled workers (Card & DiNardo, 2002).

Literature review
The impact of technology on the demand is obvious: absolute figures have increased, and despite the relatively large increase in real wages, this has not been reflected in the unemployment rate. The conjecture that technology "destroys" (or "steals") jobs is in the eyes of society more, because "winners" of introducing new technologies are not seen as much as those "defeated" (Manning, 2004). Literature dealing with SBTC sometimes shows the problems of higher inequality (see Acemoglu, 2002), but this is contradicted by Card & DiNardo (2002), according to whom the SBTC cannot explain any other dimensions of wage inequalities such as age or race. Manning (2004) pointed to the neglected fact: "average job" is improving. Weiss (2008), on the other hand, said SBTC does not necessarily deepen wage inequality. The reason is that wages do not depend on natural productivity only but also on money productivity. Consumer preferences (omitted in this area of research) stand behind the growth in demand, which is reflected in wages (even without growth in natural productivity).
Contributions about the issue of job polarization have to deal with methodological problems that arise from the identification of this phenomenon. The first problem is finding a suitable sorting criterion which will allow dividing employees into high-, middle-and lowskilled. Most often, the division is determined by individual occupations, which are classified according to the type of work performed and the skill level Another problem is the (un)availability of the data sources needed to identify the polarization. The most data (and at the same time the best available) are provided by the US labor market. This is one of the reasons why attention turned there first, and why the US labor market is the best documented. Job polarization in US has been proven since the 1980s ( What can explain job polarization? Autor et al. (2003) came up with an explanation, which was called "ALM hypothesis". Later it was renamed by Goos et al. (2014) as routine-biased technological change ("RBTC"). This title is self-explanatory. SBTC explains the changes based on skills, RBTC based on tasks. Autor et al. (2003) divided occupations by tasks into the routine, non-routine, manual, and cognitive ones. Workers who perform routine tasks can be replaced by a new technology: a machine or an algorithm (technology is a substitute for them); while workers who perform non-routine tasks will face an increasing demand as they are complementary to new technologies. 4 Thus, in general: computerization does not replace skills but tasks. Moreover, if a routine profession is well defined, then it is also possible to replace a middle-skilled worker with a less qualified worker (which is usually the case of offshoring). The RBTC is an explanation widely accepted among economists (Autor et al., 2003, Manning, 2004, Goos & Manning, 2007. It is certainly not a rejection of the SBTC theory as such, skillupgrading across sectors and professions has been demonstrated by many studies (Fernandez-Macias, 2012). The RBTC only complements the SBTC by explaining the polarization. Additionally, it is important to remember that not all the tasks (middle-or lowskilled) can be replaced by new technologies. There are still constraints resulting from imperfect technologies in specific areas or from economic constraints: it is not always worthwhile to introduce new sophisticated technologies.
Offshoring is another explanatory factor of job polarization (Goos, et al., 2009(Goos, et al., , 2014Crinò, 2009;Oldenski 2014). In general, job polarization is due to technological change (computerization) and globalization (offshoring). We can add one more reason, which is not documented as much as the impact of ICT and offshoring, but is relevant. Mazzolari & Ragusa (2014) documented that the rise in demand for low-skilled workers is partly due to the higher consumption of high-skilled workers whose wages grow faster (see Acemoglu & Autor, 2011), and can therefore demand a larger amount of low-skilled (personal) services. In other words, growing wealth of high-skilled workers pulls up low-skilled workers.
Polarization is not a domain of only a few industries, where the greatest proportion of routine jobs could be expected. Autor et al. (2015) showed that job polarization is reflected both within manufacturing and non-manufacturing sectors. This implies a structural change across the whole labor market (Goos et al., 2009; Bárány & Siegel 2018, A). That is why we are focusing on the whole economy divided into individual industries. Finally, given that wages are affected by the demand for workers, it is not surprising that polarization is also reflected in wages. Goos & Manning (2007) showed that when occupations are divided by wage distribution, we can find routine jobs in the middle, while non-routine occupations are at the edges of wage distribution.

Data and Methodology
Given the research question, it is necessary to define the relationship between skill levels and the highest level of education attained. Manning (2004) considered those who are more educated to be skilled workers. Similarly, Crinò (2009) and Krueger (1993) considered those more educated to be high-skilled workers. Mutual interchangeability of skill levels and level of education is common in literature, see Michaels et al. (2014). Thus, for the division of skills, the highest level of education attained is used. 5 Table 1 contains information about the highest level of education achieved and the level of skills. In literature, we most often encounter a division of skills into three categories; therefore, three categories are used also in this study. 6 If workers are rewarded on the basis of their productivity, changes in relative productivity must be reflected in changes in their relative wages (Acemoglu, 2002). Therefore, the relative changes in the volume of wages for high-, middle-and low-skilled workers across sectors are also reported in the results.
The research question will be answered based on the observation of changes in the proportion of employees and their highest level of education attained. Another measure will be the relative changes in the volume of wages divided by the highest level of education attained. This contribution contains descriptive methods 8 and a correlation analysis is used to measure the dependence between the relative change in the number of high-, middle-and low-skilled workers and the change in their relative wages.
The main data source is the EU KLEMS project database. 7 The EU KLEMS project includes a database containing information on productivity, technological change, economic growth, capital formation and others. All data are at NACE (Rev. 2) level of industries. The individual indicators differ in the length of their time series; the longest ranges up to 1970. The period available for the changes in education and wages used there covers the years 2008-2015. 8 Over the years, twenty-eight EU countries have been investigated. The database provides data for each year and country by gender, age (15)(16)(17)(18)(19)(20)(21)(22)(23)(24)(25)(26)(27)(28)(29)(30)(31)(32)(33)(34)(35)(36)(37)(38)(39)(40)(41)(42)(43)(44)(45)(46)(47)(48)(49), 50 years, and over) and the highest education attained for eighteen industries according to NACE classification (sector AR). This article uses abbreviations for each industry, the full names are referenced here: The results chapter includes two terms that need to be explained: trend improvement and trend worsening. The structure of each industry is composed of a high-, middle-and loweducated workers share. For example, in the Czech Republic in 2008, middle-educated workers made up the largest share in the ICT sector. Starting in 2012 (until the end of the reporting period of 2015) a change occurredthe largest proportion was made up of higheducated workers. This change in trend is "trend improvement"; the level of education in the industry improved. "Trend worsening" is the opposite.

Changes in the structure of employment based on education at sectoral level in all EU countries
First, attention is paid to changes in the structure of education level across industries. The results are shown in Table 2 below. Trend improvement/worsening means that there is an improvement/worsening in the trend which will take effect at least in two consecutive periods (see paragraph above). For example: France had the largest share of middle-skilled workers in the "D" industry during 2008-2012, but during the years 2013-2015 there was a change in the proportion towards high-skilled workers; that was an improvement in the trend. The same applies to other countries.   The results of the changes in the structure of the labor market at the level of education correspond to the SBTC hypothesisthey imply that there is primarily a rise in high-skilled workers at the expense of middle-skilled and low-skilled workers. Improvements in most industries (five in total) can be seen in Malta. A relatively large improvement was found for Denmark, Ireland, and the Netherlands (in three industries). On the contrary, there was no improvement in Belgium, Greece, Spain, and the UK. Except for these countries, there was typically an improvement in one (ten countries) or two sectors (nine countries). Trend worsening only occurred in Cyprus, Denmark, and Hungary. Altogether, there were forty-two positive trends in the period under review, while a negative trend was seen only three times. Table 3 summarizes the trend improvement/worsening at the industrial level. The trend in the "J" industry (ICT) most frequently improved (in seven countries), reflecting the development of this sector in the examined period. There was no improvement in the "F", "G", "H", and "P" industries. It is not difficult to explain the absence of trend improvement in the "P" (Education) industry. This industry had the largest proportion of high-skilled employees at the beginning of the examined period, therefore no improvement was expected. A trend worsening occurred in two sectors only: Electricity & Gas; and Arts, Entertainment and Recreation. Industries that are stagnant or have a tendency to worsen (in the case of Electricity & Gas) are considered to be, according to Amoroso & MoncadaPaterno-Castello (2018), middle-skilled industries. Their criterion is proportion of high-, middle-, and low-skilled workers according to the ISCO classification in each industry. 9 Table 3 The results presented in Table 4 below can be divided into three parts. The first part includes a relative change in the proportion of high-, middle-and low-skilled workers. This is the sum of all 18 NACE sectors for each EU country. In the second part, there is a relative change in the volume of wages of high-, middle-and low-skilled workers. Again, this is the sum of all industries. Finally, in the third part of the table, a correlation coefficient is used to measure the strength of dependence between the first and the second parts of the table: between the relative change in the proportion of skilled workers and the relative change in their wages.  9 Evaluation of each industry is based on the assumption of a homogeneous structure of occupations within each sector. If similar simplifications were to be used in this paper, we could say that the "P" industry (Education) is a high-skill industry because it contains the highest proportion of highly educated and hence high-skilled workers. In the Czech Republic (second line of Table 4), the share of high-skilled workers increased by 8 percentage points between 2008 and 2015. This increase occurred at the expense of middleskilled workers (-6 p.p.) and low-skilled workers (-2 p.p.). The share of wages of highskilled workers in all wages increased by 2 p.p., while the share of wages by middle-skilled workers did not change and the share of low-skilled decreased by 2 p.p. The correlation coefficient, which indicates the dependence between the change in the number of workers according to their attained level of education and the change in their wages, indicates a value of almost 0.7, but is not statistically significant.
We can see that the growing proportion of high-skilled workers was a trend in all countries. Typically, figures increased between 3-10 % between 2008 and 2015. Austria is significantly out of this range (15%). For middle-and low-skilled workers, the trend is a reduction in proportion. It should be noted that this loss can be caused by: death, migration, and retirement. The trend is more severely violated only in the cases of Portugal, Denmark, and Finland. An increasing trend in wage volumes for high-skilled workers is in line with the rising demand for them (SBTC hypothesis). However, there are countries where the drop in the volume of lowskilled workers is lower than or equal to that of middle-skilled workers. This, due to the theoretical assumptions about the relationship between labor demand and wages, implies that the demand for middle-skilled workers drops faster than the demand for low-skilled workers (this could point to job polarization). This is happening in eight countries: Croatia, Latvia, Lithuania, Hungary, Austria, Poland, Romania, and Sweden. Evolution in the remaining twenty countries corresponds to the requirements of skillupgrading and its linkage to wages. The correlation coefficient is significant only in one third countriesthat is, the relationship between the change in proportion of high-, middle-and low-skilled workers and the volume of their wages in all wages is statistically significant only in nine EU countries.

Changes in productivity and wages caused by the change in the number of higheducated workers
The previous results have clearly demonstrated an increase in high-educated workers across all EU countries. Naturally, the greater proportion of more skilled (educated) workersskillupgradingraises the question of how much this has translated into productivity growth. In this section, we will narrow the research area to the Czech Republic and the surrounding countries (Germany, Austria, Slovakia, and Poland). Again, we will use the NACE classification of the industries, but now not every sector will be separate. We will use Eurostat's approach where some industries with similar characteristics are combined. Four sectors of B-E are linked together, but the C-industry appears also separately, as it plays a dominant role in many countries (such as the Czech Republic). In addition, there are three GI industries (all classified as less knowledge intensive services), two M-N industries and three O-Q industries (knowledge intensive services). Productivity is expressed as value added per employee and its relative changes are observed (Bárány & Siegel, 2018, B;Ngai & Petrongolo, 2017). This indicator is also provided within the EU KLEMS database; its construction and comparison with other indicators is discussed in O'Mahony & Timmer (2009). First, Table 5 below provides the change in the number of high-educated employees in percentage points. Then, changes in productivity are shown in Table 6. Finally, Table 7 contains changes in wages of high-educated workers. It is observed how the wages of higheducated workers have changed compared to all wages in the economy (the sum of loweducated, middle-educated and high-educated workers). A-R are industry abbreviations; their full names are given in the methodology. Table 5 shows that the rise in high-educated workers is at the level of all industries in all countries. For example, in Austria, which is in the first line, the number of high-educated workers rose in all industries. In industry A, their share grew by 7.8 p. p., in industry J by 27.65 p. p., etc. There are only six minor exemptions that affect three countries -Germany (industry F and O-Q), Denmark (A, F, L), and Latvia (L). There was a decrease in the number of higheducated workers in these industries. A negative sign means that either (a) the level of education may have increased, or (b) the employee was moved to another industry, or (c) to another country. As these are high-educated workers, increasing the level of education is out of the question, so moving to another industry or another country is a possible explanation. Transfers to other countries should be of a marginal nature, so transfers between sectors are the most likely explanation.
The the biggest changes in the number of high-educated workers occurred in Austriavery significant changes amounting to 25-27 p. p. were carried out in sectors that are knowledge demanding: J, K, L. The Czech Republic as well as Poland also achieved relatively large growth values, albeit not as in Austria. The growth of high-educated workers in Slovakia was a bit lower, but it was still higher than in the case of Germany, which in terms of growth values ranked the last within the neighboring countries. The common denominator of these countries was the largest increase in high-educated workers, typically in the J, K and O-Q sectors (all considered knowledge intensive services). This reflects the EUwide result, where higheducated workers grew the most in the same industries, i.e. industries that are non-industrial in nature and rank among knowledge intensive services. Table 6 below shows the percentage change in productivity in each industry. Change in productivity is represented by a change in value added per employee. The evaluation of the results from Table 6 assumes that they are compared with Table 5. As a prerequisite, we could refer to a situation where the number of high-educated workers is increasing in the industry and this will translate into higher productivity (higher value added per employee). For example, in Austria, which is in the first line, productivity in industry A decreased by 15.76 p. p. and in B-E by 0.25 p. p. In other industries, the expected productivity growth has already occurred (expected due to the increase in the number of high-educated workers in all industries). Looking at both tables, we can see that this assumption is violated in some countries and industries. The fastest growth of productivity within the EU can be found in manufacturing (C), financial intermediation (K), and construction (F). The first two are the only sectors where, without exception, the growth of high-educated workers is accompanied by growth in value added per employee.
The assumption of a direct relationship between the change of high-educated workers and the change in productivity applies only to Poland. For the other countries, there is always at least some exception. Anomalies can be especially found in industries A, J, and MN.
We can find a situation where the decline of high-educated workers led to growth in productivity. In industry J (ICT), a relatively large increase in high-educated workers resulted in growth in productivity, but not everywhere (it dropped in Slovakia) and not so much across the EU as in other industries: industry F (construction) grew four times more and industry C (manufacturing) even five times more. While the M-N industry manifested a relatively large increase in high-educated workers, productivity gains were not so great compared to other industries. Even in the Czech Republic and Slovakia, productivity fell despite the rise in higheducated workers. M-N industry grew as the second slowest throughout the EU, albeit that includes M (professional, scientific and technical activities). A-R are industry abbreviations; their full names are given in the methodology. Table 7 above shows the variability of the volume of wages of high-educated workers in the wages of all workers. The first line, which represents Austria, shows that the wage volume of high-educated workers increased in all industries except R, where it fell by almost 2% at the expense of middle-and low-educated workers. This almost describes the EU average (sixth line) trend, which implies volume growth across all industries in the order of percentage, except for B-E industries (over 26%) and G-I (over 13%). Other countries have different results and the decline in the share of high-educated workers in the wages of all workers, such as Germany, is not an exception.

Conclusion
This paper contributes to the debate on the current changes in labor markets; specifically, it looks at changes in the structure of employment based on the education criterion in the EU countries between 2008 and 2015. The labor market can be examined from different perspectives. One option is to observe how employment structure is changing in terms of skill levels and educational attainment. Some of the existing literature puts these aspects together, making it possible to consider more educated as more skilled (Acemoglu, 1998;Krueger, 1993;Manning, 2004;Crinò, 2009).
Thematic contributions focusing on changes in structure of employment show that the demand for middle-skilled workers is declining, leading to skill-polarization (see Autor et al., 2003;Goos & Manning, 2007;Goos et al., 2014 and other papers which identify job polarization), while initial expectations assumed that the employee structure would evolve in favor of more educated employees, leading to skill-upgrading (SBTC hypothesis)see Acemoglu (2002), Card and DiNardo (2002) and Manning (2004). This paper turned attention to the employment structure in all eighteen industries according to NACE. The results imply a growing proportion of more educated workers in twenty-five EU countries. Only in three countries, there was no improvement in the "education trend" over the examined period in any industry. In terms of frequency, the greatest improvement was seen in the ICT sector, but improvements are more or less spread across all industries. The worsening education trend took place marginally within two countries and two industries. In general, the potential of the labor market is growing, providing support for the SBTC hypothesis, which assumes a rise in demand for people with a higher skill level. In addition, changes in the proportion of wage volumes were examined for each country. The sorting criterion was again the level of skills which was determined by the level of education attained. Developments in EU countries correspond to the assumptions of skillupgrading and its linkage to wages.
The results implying skill-upgrading rather than skill-polarization have raised the question of whether a higher skill level (higher level of education attainment) affected productivity, expressed as value added per worker. Here the area was narrowed from the EU28 to the Czech Republic and its surrounding countries. First, it was necessary to observe the relative changes of high-educated workers in all industries at country level. Poland and Austria recorded slightly higher proportions of high-educated workers than the Czech Republic, Germany, and Slovakia. Then, changes in productivity were observed and compared with the change in the number of high-educated workers. Additionally, other results have not clearly shown a link between the change in productivity of high-educated workers and their wages (vs. Acemoglu, 2002). For example, in Germany the productivity of high-educated workers increased in all industries except Agriculture (A), but the wage volume of high-educated workers (of all employees in the economy) decreased in all industries.
This contribution also shows how the composition of individual industries changed. For example, the results confirm that in industries Wholesale and Retail Trade, Transport and Storage, Hotels and Catering (industries considered less knowledge intensive services), the structure changed less in favor of more educated employees than in Public Administration, Health and Social Work (industries considered knowledge intensive services).
The growing proportions of more educated employees open up space for further research because this fact may not reflect the reality that people with a higher level of education perform more demanding tasks required by this higher education. Higher education is not a guarantee of the profession that requires it; these people are only more likely to have an adequate job corresponding to their skills. In practice, it is not an exception to meet people with higher education, yet they do a profession that does not require any professional education. The literature dealing with this phenomenon, called "overqualification" 10 , is very limited. For methodological problems with its measurement, see Erdogan et al. (2011).