Unconditional and Conditional Wage Polarization in Europe∗ Riccardo Massari

Paolo Naticchioni

Giuseppe Ragusa

Sapienza University of Rome

University of Cassino and IZA

Luiss University

[email protected]

[email protected]

[email protected]

Draft version, October 5, 2012

Abstract The US labor market has become increasingly polarized both in terms of jobs and wages, and the routinization explanation is well established for these trends. Recent papers have found job polarization patterns also in Europe, while few evidence is available for wages. The goal of the paper is to investigate the dynamics on unconditional and conditional -on technologywages in Europe, using industry (EU KLEMS) and individual data (ECHP and EU-SILC). As for unconditional wages, both at the industry and individual level data there are no wage polarization trends at work. For the conditional impact of the routinization explanation, at the industry level we investigate the impact of ICT intensity on wages and hours worked by three skill groups. Our analysis does not provide evidence supporting the polarization of wages, while we detect job polarization trends. Interestingly, at the individual level we have some mild evidence in favor of polarization, due to the coefficient effect of service tasks, for the lower tail of the distribution, and of abstract task for the upper tail of the distribution. Institutions plays instead a role in explaining the increase in inequality in the lower tail of the distribution. JEL Classification: J3, J5 Keywords: wage inequality, polarization, occupational tasks, offshoring, RIF-regressions.

∗

We thanks Carlo Dell’Aringa, Wen Hao Chen, Nicole Fortin, Thomas Lemieux, Marco Leonardi, Elisabetta Magnani, Raul

Ramos, Giovanni Sulis, for their suggestions, and the participants to seminars held at Cattolica University (MILLS Seminar, Milan), OFCE-Science Po (POLHIA meeting, Paris), CeLEG-Luiss (Rome), Padova, ECINEQ conference (Catania, 2011), AIEL conference (Milan, 2011), SIE conference (Rome), Royal Economic Society (Cambridge, 2012), Barcelona (UB), Cagliari, Napoli Partenope, Rome (La Sapienza).

1

1

Introduction

Autor et al. (2006) have shown that the US labor market has become increasingly polarized both in terms of occupations and wage distributions. The routinization hypothesis has been advocated for explaining this empirical evidence (Autor et al., 2003): job polarization occurs because of information and communication technologies complementing the analytical tasks performed by highly educated workers and substituting for routine tasks generally performed by middle educated workers. Goos et al. (2009) find job polarization patterns in Europe as well, and Goos et al. (2010) investigate the effects of technology, globalization, institutions and product demand effects on the employment dynamics for different occupations. Their results suggest that in Europe the routinization hypothesis is the most important factor behind the observed shifts in employment structure. Michaels et al. (2010) use EU KLEMS industry level data to study whether the distribution of wage bill shares in European countries has polarized and whether this is due to information and communication technology. Their findings suggest the routinization explanation applies. In a sense, Michaels et al. (2010) investigate a conditional impact of technology on wage polarization, without showing if actually unconditional wages have polarized over time in Europe. The goal of this paper is to add to the empirical evidence on wage polarization in Europe. First, we analyze whether the unconditional wages get polarized over time in Europe. Second, we analyze the conditional impact of technology on wages, using proxies for technological change, in such a way testing the routinization explanations. We make use of both industry and individual level data. At the industry level we use the EU KLEMS database, from 1980 to 2005, which contains data on value added, labor, capital, skills and ICT for various industries in OECD countries. In this paper we have to focus on nine countries for which the variables of interest are available (Austria, Denmark, Finland, France, Germany, Italy, Netherlands, Spain, the UK). At the individual level, we created a new data source harmonizing the European Community Household Panel (ECHP) and the European Union Statistics on Income and Living Conditions (EU-SILC). We make use of two different samples, the first one concerns the sample of countries for which it is available the gross current hourly wage (AT, ES, GR, IE, IT, PT, UK), the second concerns a wider set of countries (AT, BE, DK, ES, FI, FR, GR, IE, IT, LU, PT, UK) for which the gross yearly earnings of the previous year is available. As far as the dynamics of unconditional wages is concerned, we show that both using industry

2

and individual data unconditional wages do not polarize over time. For the conditional impact of technological proxies, the evidence is more mixed. Using EU KLEMS industry data, we investigate the impact of new technology on relative wages between high and medium skilled, and between medium and low skilled. We point out that changes in ICT intensity has impacted the labor market through a polarizing effect in jobs. The evidence on the relation between wage polarization and ICT is much weaker. We find evidence supporting ICT as having had an effect on the high-skilled (HS)-medium-skilled (MS) relative wages, but no effect is found for the medium-skilled (MS)-low-skilled (LS) relative wages. We also find that the financial sector has played an important role in Europe: without this sector there is no evidence of European wages responding to changes in ICT intensity. Apart from some limitations in the EU KLEMS data, such as the not always homogeneous definition of educational levels across countries, it is worth stressing that EU KLEMS provides a proxy for technological change at the industrial level, while the recent literature on this topic (Autor and Dorn, 2012; Firpo et al., 2011) has shown that technological progress can affect also the within sector dynamics, mainly with respect to different tasks in different occupations. In a sense, the industry level data captures the between sector impact of technological change, while in the individual level analysis the focus will be on the between occupation technological impact, even within the same sector (and/or firm). Further, using individual level data it is possible to control for individual observed heterogeneity, and it allows investigating the whole wage distribution instead of limiting the analysis on three educational groups. For all these reasons, we argue that the industry and individual data analysis provide different interesting insights on polarization issues in Europe. For conditional polarization using individual data, we augment the ECHP and EU-SILC data with occupational task measures derived from Goos et al. (2009). These measures are intended to proxy for jobs’ technological characteristics such as the level of routinization and the extent to which jobs are offshorable (Goos et al., 2010). We exploit individual level data by means of a recent decomposition methodology (Fortin et al., 2010), which allows identifying a detailed composition and wage structure effects along the whole wage distribution. In the sample of gross current hourly wages, we find some evidence, although weak, in favor of a polarizing impact of technology. This evidence is mainly due to the wage structure effect, and in particular to the service task for the lower tail of the distribution (Autor and Dorn, 2012; Mazzolari and Ragusa, 2012) and to the abstract task for the upper tail of the distribution (Firpo et al., 2011). The impact of offshoring is instead constant along the wage distribution, consistently with 3

Goos et al. (2010). The other components of the wage structure effect (demographic, institutions, education), as well as the detailed composition effects, do not entail a polarizing effect on wages. Similar findings are confirmed when using the gross yearly earnings, even if the U-shape technological impact is less pronounced. In particular, also in this case service is the driving force for the lower tail of the wage distribution, while abstract drives the upper tail. Among the composition components, had only the composition of institutions changed over time, i.e. increases in part time and temporary contracts, the wages at the 10th percentile would have strongly reduced, entailing a not negligible impact on the increase in the 50-10 ratio. Finally, it is interesting to note that according to the individual level analysis, education in Europe does not exert a positive impact on wage inequality and it is not a driver of wage polarization. This is at odds with the literature for the US (Firpo et al., 2011), while being consistent with OECD (2011). The paper is structured as follow. Section 2 presents the analysis using aggregate industry data, while section 3 focuses on the individual data analysis. Section 4 concludes.

2

Macro evidence

In this section we present aggregate evidence on the polarization of the European labor market. The analysis is based on the EU KLEMS Growth and Productivity Accounts. The EU KLEMS database contains industry–level measures of output, inputs, and productivity for 23 European countries, Japan and the US for the period from 1970 to 2005. The database harmonizes data provided from each country’s National Statistical Office. O’Mahony and Timmer (2009) provide a detailed explanation of the construction of the dataset. For a selected number of countries—Austria, Denmark, Finland, France, Germany, Italy, Japan, the Netherlands, Spain, the UK, and the US—the database collects industry–level data on wages and hours worked by skill levels: high (HS), medium (MS), and low skilled (LS). Workers are assigned to these skill levels according to their education attainment. Key to the analysis here, EU KLEMS has industry level data on Information and Communication Technology (ICT) capital compensation; we use this variable to construct a measure of industry intensity in new technologies. There are at the time of writing four different releases of the EU KLEMS database: March 2007, March 2008, November 2009, and March 2011. We use the most recent release (March 2011) for the industry input data, and the March 2008 for the skill-level labor data. This choice is due to the fact that skill-level labor data was not updated in the last two releases. Differences in the March 2008 and March 2011 are modest and the findings of this section are not sensitive to the 4

use of either one of the dataset. On the contrary, the March 2008 and the March 2007 releases diverge in a substantial way. As such, some of the conclusions of Michaels et al. (2010)—who uses the March 2007 release—do not survive the data upgrading. While EU KLEMS provides unique information on industry–level economic variables and their dynamic over time, the database has some inconsistencies that should be kept in mind when interpreting the results. First, the criterion with which skill level groups are identified varies across the eleven countries. While the definition of high skilled group is basically homogeneous among countries—high skilled are those with a college degree—the medium and the low skilled groups are less uniformly defined. This results in a problem with some European countries having an unreasonably low fraction of low-skilled workers. Second, the number of industries for which skill composition variables are available varies across countries. To obviate this problem, we follow Michaels et al. (2010) and aggregate industries to the lowest possible level of aggregation for which all skill-level variables are present. Also, in EU KLEMS capital compensation measures are built using the perpetual inventory method from the underlying investment flow data for several types of capital. In this approach, the price of capital services is defined as a residual and in practice it can be negative. When this happens—and it does frequently in the agriculture industry—we set the price of capital to zero. [Figure 1 about here.] [Figure 2 about here.] Figure 1 and Figure 2 plot—for each European country in our sample—the by-industry 19802005 changes of our measure of ICT intensity: ICT capital compensation over value added (ICT/VA),1 distinguishing tradeable from non-tradeable industries.2 In the final sample, each country has a minimum of 14 industries, with Italy, Denmark, US, and Japan having data for 27, and Spain for 19. The number of tradable industries is limited in many countries: Austria, Finland and Netherlands have data for only three tradeable industries; Germany, UK, and France for four. The highest growth in ICT/VA took place in non-tradable industries. In particular, in six of the 9 countries 1

The choice of value added as scaling factor is arbitrary and other variables related to the whole size of the

economy could be potentially used. We use value added to be consistent with Michaels et al. (2010). 2 Tradeable industries are: electrical and optical equipment; pulp, paper, paper products, printing and publishing; transport equipment; machinery; chemical, rubber, plastic products; wood and products of wood and cork; basic metals and fabricated metal products; non-metallic mineral products; textiles, textile products, leather and footwear; food products; agriculture, hunting, forestry, and fishing.

5

(Denmark, Spain, Finland, France, Germany, and Netherlands) the highest growth occurred in the financial industry. In Finland, the ICT/VA grew for the financial industry by 25 percentage points. In Austria and Denmark by 15 percentage points. In the remaining countries, the Financial industry registered more modest increases, but always disproportionate with respect to the other industries with the exception of the postal and telecommunication services. [Figure 3 about here.] Figure 3 shows the 2005-1980 changes of ICT over value added aggregated for our sample of European countries. As it is to be expected, also at this level of aggregation the industries with the highest increase in our measure of ICT intensity are non-tradeable industries: financial and postal. [Table 1 about here.] Table 1 shows summary statistics for the 1980 level and the 1980-2005 change of the key variables. The variables are wage bill share, relative hours worked, and real wages, all broken down by skill levels; ICT and non-ICT capital over value added, and ICT capital over total capital. All monetary quantities are expressed in real US dollars.3 Finland is the country with the highest increase in the wage bill accruing to the high skilled. Countries with similar rate of growth are UK, USA, JPN, and Spain. The wage bill accruing to the medium skilled declined only in the US and in Germany. In all countries, hours worked over total hours increased for high skilled and decreased for low skilled. The US is the only country that experienced a drop in the share of medium skill hours worked. ICT capital intensity increased in all countries although there are marked differences on the rate of growth. The country with the largest increase is the USA, while Germany, France and Italy are the countries with the smallest change in ICT intensity over the 1980-2005 period. [Figure 4 about here.] Using our data, we can try to asses to which extent changes in technology, in this case changes in our measure of technology—ICT compensation over value added, have had any role in polarizing 3

The table makes clear that the issues with the skill identification criterion adopted by EU KLEMS may be rather

severe. For the year 1980 Italy has an excessively small low skilled wage bill share (7.70%) which drops to 0.28% in 2005. Both the level and the change over the 1980-2005 period are not in line with those that can be inferred from other Italian dataset. EUklems’ definition used for low skilled for Italy is “No formal qualification” and the first diploma that can be earned in this country is the elementary school one, which can be obtained only after 5 years of education. The extremely low level in he 1980 and the large drop in the 1980-2005 could be explained by this.

6

the European job market. Figures 4 draws 1980-2005 changes in ICT/VA against changes in the three skill groups wage bill share for all European countries. The wage bill share of high skilled has on average increased more in industries that have experienced higher changes in ICT/VA over the 1980-2005 period. There is instead a negative correlation between changes in wage bill share of medium skilled and changes in ICT/VA. For low skilled, changes in wage bill share do not to have any correlation with changes in ICT/VA. Remarkably, the correlation between wage bill share and ICT/VA is disproportionately influenced by the financial and postal sectors both for high skilled and medium skilled. At the country level, the correlations between wage bill shares of the three education groups and ICT/VA follow a similar pattern. As highlighted in Figure 5, the correlation between changes in the wage bill share of high skilled and changes in ICT/VA are positive in almost country with the exception of Italy; for medium skilled (Figure 6) the correlation is negative in all countries, although very small in Italy, Germany and the UK. When looking at the correlations for the low skilled groups, there is more heterogeneity (Figure 7). As in the aggregate case, the financial and postal industry are outlying observations which have a strong effect on the sign of the correlations. [Figure 5 about here.] [Figure 6 about here.] [Figure 7 about here.] This graphical investigation suggests that there exists a relationship between changes in ICT investment and changes in the wage bill share that is broadly in line with the routinization hypothesis. To further shed light on how the behavior of ICT investment has impacted the labor market outcomes of different skill groups we turn to a regression analysis. In the first set of regressions, the dependent variables are 1980-2005 changes in high skilled, medium skilled, and low skilled wage bill share, respectively. These variables are regressed on the 1980-2005 changes of ICT/VA a set of controls: the changes of non ICT capital over value added, and the log change in value added over the same period. Regression results are reported in Table 2. Column (1)-(3) report baseline OLS estimates, which also includes country fixed effects to capture country specific trends. The signs of the coefficients on ∆ICT /V A are positive and statistically significant for high skilled and negative and statistically significant for medium skilled. In the case in which the dependent variable is the change in the low skilled wage bill share, the coefficient on ∆ICT /V A is positively estimated

7

but statistically not different from zero. To gauge the economic significance of the estimated coefficients, consider that a one standard deviation increase in ∆ICT /V A explain about 18% of a standard deviation of high skilled wage bill share and 9% of a standard deviation of medium skilled wage bill share. Even from an economic significance point of view, the coefficient on ∆ICT /V A in the case of low skilled wage bill is insignificant—a one standard deviation increase in ∆ICT /V A explain less than 1% of a standard deviation of this group wage bill share. [Table 2 about here.] In columns (4)-(6) of Table 2, we re-estimate the same specification omitting the observations relative to the financial and the postal industries. Once these two industries are removed, the magnitude of coefficients on ∆ICT /V A is smaller and none of these coefficients is statistically significant. This suggests that much of the wage bill share polarization could be due to the effect of ICT in few technology intensive industries. Columns (7)-(12) re-estimates the models in columns (1)-(6) by controlling for the 1980 level of ICT/VA, NICT/VA, and log VA. This robustness check do not dramatically changes the coefficients of interest. Next we study the impact of ICT on the log relative wage bill share and its components. Note that log changes in relative bill share can be decomposed into log relative changes in hours worked and log relative changes in wages, as follows:       WHS LHS BSHS = ∆ log + ∆ log , ∆ log BSM S WM S LM S       WM S LM S BSM S = ∆ log + ∆ log . ∆ log BSLS WLS LLS Using this decomposition, we regress the log relative wage bill share, the log relative wage change, and the log relative hours change on ICT/VA, NICT/VA, and log value added, including country dummies. Table 3-Panel A presents OLS regression results. The high skilled relative to the medium skilled wage bill share is estimated to be positively correlated with ICT/VA changes. On the other hand, the medium skilled relative to the low skilled wage bill share is negatively correlated with changes in ICT/VA, but the coefficient is very small and statistically insignificant. Interestingly, when we exclude the financial and the postal sectors, results presented in Table 3-Panel B show that the magnitude of the estimated coefficient on high skilled to medium skilled wage bill share decreases rather substantially and it also becomes statistically insignificant. As we decompose wage bill shares into hours and wages, we find that changes in high to medium skilled wage bill share tend to be almost totally explained by changes in relative hours worked. For 8

the medium to low skilled case, the effect of ICT on wage bill share is small and insignificant and thus interpreting the decomposition is of limited interest, if not for saying that there is no evidence of technology driven wage polarization for the countries considered here.4 [Table 3 about here.] We also estimate the same specifications of Table 3 by two-stage least squares, instrumenting changes in ICT over value added with the 1980 level of ICT over value added in the US. The IV results are presented in Table 4. The results are similar to those obtained by OLS. The main difference with Table 3 is that the high to medium wage bill share is estimated to be positively affected by ICT even when the financial and the postal industries are excluded from the regression. [Table 4 about here.] We interpret the results of this section as suggesting that, as for the US, the polarizing forces of technology may have had an impact on the European labor market. In particular, technology may have affected skill employment composition, but little wages. We do not find any evidence indicating that wages have responded as they have in the US. We are very cautious with our statements for a series of reason. We have found that the EU KLEMS data have severe limitations for analyzing the relationship between technology and labor market using industry data. Even disregarding these concerns, we only make a course attempt at dealing with all the potential endogeneity using instrumental variables. Furthermore, by only using aggregate data it is not possible to control for individual observed heterogeneity, and also to investigate changes of all percentiles of the skill distribution. We should keep in mind that we used only a proxy for technological change at the industry level, i.e. ICT compensation. Recent literature on this topic (Autor and Dorn, 2012; Firpo et al., 2011) has shown that technological progress can affect also the within sector dynamics, mainly with respect to different tasks in different occupations. Industry level data can only capture the between sector impact of technological change. In the next section we turn to individual level data in order to analyze between occupation impact of technological change. 4

These findings are robust to the inclusion of 1980 level of value added, ICT, and non ICT capital. In the interest

of space, we do no report the results of this robustness check here.

9

3

Micro evidence

The main novelty in the individual level data analysis of this paper is to consider Europe as a Whole. Actually for some European countries there are papers on wage inequality trends, and evidence is mixed. Actually, wage polarization have been detected only in Germany (Dustmann et al., 2008) and Portugal (Centeno and Novo, 2009), and increasing inequality has been observed for the UK (Machin, 20011). For other countries, decreasing trends in wage inequality have been observed, such as in France (Charnoz et al., 2011), Spain (Izquierdo and Lacuesta, 2006) and Italy (Naticchioni et al., 2008). Another novelty of the paper is to link occupational proxies for technology to inequality trends, something that it is not considered so far in the European literature on wage inequality.

3.1

Data

There is currently no single data source to study the dynamics of the wage structure in Europe. For this reason we have harmonized two different data sources, the European Community Household Panel (ECHP), and the European Income and Living Conditions (EU-SILC).5 ECHP is a longitudinal survey conducted (from 1994 to 2001) in 15 European Union Member States and carried out yearly under Eurostat (Statistical Office of the European Communities) coordination. EU-SILC is a yearly survey started in 2004 and includes also new EU members, and some non EU countries (Norway, Iceland and Cyprus). The two surveys shares many features, and more inportantly, it is possible to harmonize the variables of interest in the two datasets by recoding. We have considered the 1996 wave of ECHP and the 2007 wave of EU-SILC, and we have selected employees aged between 15 and 64 years. Our key variable concerns the earnings in the labour market. The more precise variable provided in the two survey is the current gross monthly wage on the current main job. It is then possible to compute the gross hourly wage multiplying the gross monthly wage by (12/52)/(hours worked per week). The main drawback of using the gross current hourly wage is that in EU-SILC it is available only for countries for which there is no data source to compute the gender pay gap: Austria, Greece, Ireland, Italy, Portugal, Spain and the UK. 5

6

This paper is not the first one that uses ECHP and EU-SILC as a single data source. Recently, Goos et al. (2009,

2010) make use of wages from ECHP and EU-SILC at an aggregated level, as a control variable to investigate job polarization trends in Europe. 6 When considering this sample we also exclude, within each country, individuals whose wage is lower than the

10

All measures of wages in the paper are expressed at 2006 prices by means of a Harmonized Consumer Price Index (HCPI). As a robustness check, we will also consider the case in which wages are adjusted by Purchasing Power Parity (PPP) deflators (base: Euro 15 = 1), to take into account different costs of living, since we investigate inequality trends in supranational entities such as Europe (Milanovic, 2006; Brandolini, 2007). In panel A of Table 5 we report the descriptive statistics for the log of current gross hourly wage, for the seven countries available. [Table 5 about here.] On average, the current gross hourly wage increases on average by 14.6%. Switching to distributional consideration, the first decile slightly increases over time (2%), while the median and the 90th percentile increase by 13.3% and 26.6% respectively. This means that wage inequality raised from 1996 to 2007, i.e. the 90-10 ratio increased by 24.6%. Further, no patterns of wage polarization are at work in Europe, since the increase in the 90-10 index can be equally divided in an increase in the 90-50 and 50-10 indexes, by 13.3% and 11.3% respectively. Panel B of Table 5 report the descriptive statistics for the (log) current gross hourly wage in PPP. The trends are quite similar, with a lower increase in inequality (13.5%) equally divided in increases in the 90-50 and 50-10 indexes (6.8% and 6.7% respectively). Since current gross earnings are available only for a subset of countries, we also consider another variable, the yearly gross earnings related to the year before the interview. The pros for using this variable is that we can cover a much wider set of countries (AT, BE, DE, DK, ES, FI, FR, GR, IE, IT, LU, PT, UK), except Sweden and Netherland for which some of the covariates are not available. Further, we exclude Germany from the sample, since several papers (Hauser, 2008; Frick and Krell, 2010) underlines data quality problems for EU-Silc in Germany, and in particular for what concerns inequality and poverty measures.7 The cons of using this sample is that while the earnings variable refers to the previous year (1995 and 2006), all the other variables refer to the time of the interview (usually the first part of the years 1996 and 2007). While some of these covariates are persistent, such as education, other variables can change over time, such as having a temporary or a part time job, sector, occupation. This means that some measurement error can apply. 0.5th percentile and higher than the 99.5th percentile. All data are also weighted with sample weights provided by the two surveys, multiplied by the weekly hour worked. 7 In our data we had the feeling of this problem since the 90-50 ratio decreased over time for Germany, while all the literature stresses that in the same period of time the 90-50 increased.

11

Nonetheless, we claim that even with this drawback our data represents a not negligible improvement with respect to other data sources used to investigate these issues. For instance, the data used in OECD (2011) harmonizes different data sources (surveys, administrative registers and tax records) for different countries, meaning that the sample structure, the questionnaire, the variable definitions, are different across countries. According to Atkinson (2008), the OECD data are more suited for assessing changes in earnings distributions over time than for comparing levels across countries. Our data, although not perfect, overcomes most of these data limitations. Panel C of Table 5 reports the trends for the (log) yearly gross income in the previous year. We also exclude from the sample those individuals working less than 12 months in the previous year and earning less than six thousand euros gross, to minimize the impact of differences in months worked and to disregard workers less attached to the labour market. It is interesting to note that using yearly earnings the 10th percentile decreases over time (-3.7%), while the median and the 90th percentile increase, by 11.6% and 13.1% respectively. This means that the ratio 90-10 raises by 16.8% and it is concentrated in the lower tail of the distribution, i.e. the 50-10 ratio increases by 15.4%. When considering the (log) yearly gross earnings in PPP (Panel D), similar findings are derived, with the 90-10 increasing by 12.3%, concentrated in the lower tail (8%). The main message from the four panels of Table 5 is that inequality increased in Europe in the periods considered, even though unconditional wages do not polarize over time, mainly due to the fact that the change in the 50-10 ratio is always positive instead of negative. These patterns confirm the ones observed in the macro analysis concerning the not polarizing trends of unconditional wages.8 As covariates for our analysis, we recover from the ECHP and the EU-SILC database the following variables, using harmonized definitions: gender; potential experience; education level: Primary, Secondary and Tertiary; having a Permanent vs a temporary job; having a full time vs a part time job; industry: Manufacturing, Wholesale, Restoration and Transport, Financial Intermediation and Business Activities, Public Administration, Education and Health, Others Services. Panel A of Table 6 includes the descriptive statistics for the sample for which current gross current (hourly) wages are available (AT, ES, GR, IE, IT, PT, UK). During the observed period female work participation increases pervasively, employees become, on average, more educated and slightly more experienced. The employment share in services increases, while manufacturing 8

Note that inequality trends are not much homogeneous across countries. However, in none of the country wage

polarization trends are at work.

12

decreases. Temporary and part time contracts increase as well, consistently with the reforms concerning labour market flexibility adopted in many European countries in the last decades. Similar trends are derived when considering the wider sample of countries, for which the yearly gross labour income for the previous year is available (AT, BE, DK, ES, FI, FR, GR, IE, IT, LU, PT, UK), as shown in Panel B of Table 6. [Table 6 about here.] As for the technological proxies, we rely on the recent literature stressing the importance of occupational level tasks (Autor et al., 2003; Goos and Manning, 2007; Lemieux, 2008). In particular, we consider three task measures at the occupation level, regarding Routine, Abstract and Service task intensity. Routine tasks are those that can be segmented in step-by-step rule. Non-routine tasks are split into two categories, Abstract and Service, to distinguish the cognitive skills required to complete these tasks. These measures are derived from Goos et al. (2010), that in turn uses the Occupational Information Network (ONET) database.9 There is also an increasing interest in the impact of globalization on employment structure, i.e. outsourcing of parts of the production process focused on specific occupations (Feenstra and Hanson, 1999; Grossman and Rossi-Hansberg, 2008). For this reason we make use of another task variable concerning offshorability intensity, derived from Goos et al. (2010).10 All these measures are available at the 2-digit International Standard Occupational Classification (ISCO), and we impute the corresponding value from Goos et al. (2009). Then, these variable has been normalized, as explained in Section 3.2. Before switching to the decomposition analysis, it is worth noting that our data can reproduce the job polarization patterns derived by Goos et al. (2009) using the European Labour Force Survey data. In our data we can recover the same three aggregate categories as in Goos et al. (2009): the top skilled jobs, including the 8 highest paying occupations; the medium jobs, with the 9 middling occupations, and the unskilled jobs, with the 4 lowest paying occupations. In the sample of the current hourly wages, the top skilled occupations increased by 4.6 percentage points, the middling jobs decreased by 9.4 percentage points, and the unskilled ones increased by 4.8 percentage points. 9

ONET is a comprehensive database which provides data on worker characteristics, worker requirements and gen-

eral work activities in US occupations. As in Goos et al. (2010), the underlying assumption is that job characteristics are invariant across developed countries. See Goos et al. (2010) for further details about the computation of these occupational measures. 10 Goos et al. (2010) derive this variable from the European Restructuring Monitor (ERM).

13

Similarly, in the sample of the yearly gross earnings, the share of top occupations increased by 3.8 percentage points, the medium decreased by 8.8 p.p. and the unskilled increased by 5.0 p.p..

3.2

Decomposition methodology

We are interested in identifying the role of each covariate in affecting the change in wages along the whole wage distribution. To do it, we make use of the methodology proposed by Fortin et al. (2011) (henceforth FFL), that we briefly describe in the following. Let consider the wage variable Y observed over a population of employees in two periods, T = {0, 1}. With respect to the mean, the well known Oaxaca-Blinder (henceforth OB) decomposition (Blinder, 1973; Oaxaca, 1973) allows to break down the overall difference in means, ∆µO , into two components, one related to the changes in the returns of the set of covariates, the wage structure effect, ∆µS , and the other linked to the changes in the distribution of these covariates, the composition effect, ∆µX , by simply adding and subtracting a counterfactual mean of Y . It is also possible to identify the contribution of each covariate to these two effects, the so called detailed decomposition. In order to extend this decomposition to the whole wage distribution, we make use of the FFL decomposition, which similarly to OB allows decomposing the changes over time of a distributional parameter ν into the two effects described above. In particular, we are interested in the decomposition of the change over time of percentiles of the wage distribution and of the standard inequality indexes (90-10, 90-50, 50-10). Key to this decomposition is the counterfactual earning distribution Y01 , which represents the distribution that would have prevailed under the earning structure at period T = 0, with the individual characteristics observed at period T = 1. FFL show that a OB-type decomposition can be carried out expressing ν as a linear function of the covariates by means of the so-called recentered influence function (RIF)–regression. In broad terms, the RIF of ν consists of adding back the distributional parameter to the influence function (IF):11 RIF (Y ; ν) = IF (Y ; ν) + ν The RIF–regression is simply a regression function in which the dependent variable is the RIF of the distributional parameter ν of the distribution of Y . In a linear framework the RIF–regression 11

The influence function Hampel (1974) is a statistical tool, widely used to measure the robustness of a distributional

statistic to the presence of outliers. The influence function detects the contribution (also defined as “influence”) of each observation to the distributional parameter of interest.

14

can be written as: E[RIF (Y ; ν)|X] = Xβ ν where β ν represents the marginal effect of X on ν, which can be estimated by OLS. Given these premises, the OB-type decomposition of the overall change in the distributional parameter ν between period 0 and 1 can be decomposed into two effects: ∆νO = E[X|T = 1]0 (β1ν − β0ν ) + (E[X|T = 1] − E[X|T = 0])0 β0ν | {z } | {z } ∆νS =wage structure effect

(1)

∆νX =composition effect

Similarly to the OB decomposition of the mean, these aggregate effects can be decomposed into the detailed effect of each covariate: ∆νS

=

p X

ν ν E[Xk |T = 1]0 (β1,k − β0,k )

k=1

∆νX =

p X

ν (E[Xk |T = 1] − E[Xk |T = 0])0 β0,k

k=1

However, this decomposition entails a bias, due to the fact that the linear specification holds only locally.12 FFL suggest a solution based both on the Di Nardo et al. (1996) reweighing procedure and on the RIF regression. By reweighing the distribution of X’s in period 0 to have the same distribution as in period 1 ¯ 01 and the counterfactual coefficients βˆν from it is possible to estimate the counterfactual mean X 01 ˆ (Y0 ; ν) on the reweighted sample. Then, one can estimate the reweighted the regression of RIF ˆ ν , and of the specification ˆ ν as the sum of the “pure” composition effect, ∆ composition effect ∆ X,p X,R ν ˆ error, ∆ : X,e

ˆ ν = (X ¯ −X ¯ )ˆ ¯ [ˆ ∆ γν + X γ ν − γˆ0ν ] X,R | 01 {z 0 0} | 01 01 {z } ˆν ∆ X,p

(2)

ˆν ∆ X,e

Similarly, it is possible to estimate the reweighted wage structure effect as the sum of the “pure” ˆ ν , and of the reweighting error, ∆ ˆν . wage structure effect, ∆ S,p

S,e

ˆ νS,R = X ¯ [ˆ ¯ −X ¯ )ˆ ∆ γ ν − γˆ ν ] + (X γν | 1 1{z 01} | 1 {z 01 01} ˆν ∆ S,p

12

(3)

ˆν ∆ S,e

Barsky et al. (2002) show that, in the case of the mean, the Oaxaca-Blinder decomposition is biased. FFL observe

that this bias can occur also for other distributional statistics.

15

The interpretation of the detailed wage structure decomposition raises some issues due to the arbitrary choice of the base group (Oaxaca and Ransom, 1999), both in the case of categorical variable and of continuous variable, such as task intensities. In the latter case, it is useful to ˜ normalize the continuous variable in order to reduce ambiguity in interpretation of results. Let X ˜ in the base group. As in Firpo et al. (2011) be the continuous variable and XB the value of X ˜ = 1] − XB is equal to half standard deviation of we normalize in such a way to have that E[X|T the continuous variable. In practice, we normalize the continuous variables used in the analysis (abstract, routine, service, offshore) as follows: x ˜t = xt + 0.5 · sd(X) − E(X1 ). With this normalization E(X˜0 ) = E(X0 ) + 0.5 · sd(X) − E(X1 ), and E(X˜1 ) = 0.5 · sd(X). Note that with this normalization the composition effect does not change, while the wage structure effect becomes: ν 0.5 · sd(X)[γ1ν − γ01 ].

Hence, the wage structure effect have to be interpret as the change in the dependent variable ˜ due to a half standard deviation increase of the continuous variable X. As a final remark, note that the strict exogeneity condition, usually invoked in the standard Oaxaca-Blinder decomposition, is not necessary for the identification of the decomposition terms within this framework, and can be substituted with the less severe ignorability assumption. Under this hypothesis, the expected value of residuals conditional on X need not be zero; the only requirement is that it has to be the same in the two time periods, an assumption that in our context can be considered as reasonable. Moreover, under this assumption, it is possible to give a causal interpretation to the decomposition results, in particular to the structure effect (Fortin et al., 2011).

3.3 3.3.1

Decomposition results Analysis for the decomposition of the dynamics of the gross current hourly wage

At first, we implement the decomposition methodology in the time period 1996-2007 for the sample of seven European countries (AT, ES, GR, IE, IT, PT, UK) for which the dependent variable is the gross current hourly wage.13 13

The reference group is composed by male, with 20-25 years of experience, with upper secondary education,

permanent and full time, in manufacturing. The same reference group will be used when considering the gross yearly

16

We show the results of the decomposition methodology mainly by means of a graphical representation, which is more effective in summarizing the wide set of information at each decile of the wage distribution. Consistently with the descriptives statistics, Figure 8 shows that wages increase –almost monotonically– along the distribution, entailing an increase in the 90–10 index by 24.6%. The decomposition allows separating the aggregate composition effect from the aggregate wage structure effect. It is worth noting that the wage structure effect displays a dynamics quite close to the overall wage trends, while the composition effect is much flatter along the distribution.14 [Figure 8 about here.] We then move to the detailed composition and wage structure effect, to identify the impact of each set of covariates. Similarly to Firpo et al. (2011), we construct the following six components of interest, summing the coefficients for the dummies included in each component: • education • technology (abstract, routine, service tasks) • offshoring • demographic (female and potential experience) • institutions (part time and fixed term) • sector In figure 9 we show the detailed composition effect. As first remark, all components are very close to zero, meaning that in magnitude they do not affect much the wage dynamics. As expected, technology entails an increasing impact along the wage distribution, very small in magnitude, and offshoring follows a similar pattern. It is also worth noting that the education component is the one with the strongest impact (anyway lower than 0.05 along the whole wage distribution). Rather surprisingly, the composition effect for education does not increase along the wage distribution, being indeed slightly decreasing. [Figure 9 about here.] earnings. 14 The reweighing and specification errors, computed as defined in the methodological section, are very close to zero and not statistically different to zero. We report the two errors and their standard errors in Table 7.

17

Figure 10 concerns the detailed wage structure effect. Interestingly, the technological component entails a U-shape impact on wages: had only the coefficients associated to the technology component changed over time, wages would have polarized in the lower tail of the wage distribution, since the impact on the 10th and 20th percentiles (around 8%-10%) is greater than the one at the median. The polarization of the upper part is instead less pronounced. We also decompose the technology component in the three occupational tasks (abstract, routine, service), as shown in Figure 11. Interestingly, the abstract component exerts a steep increasing impact along the wage distribution, entailing a positive impact on the increase of both the 90-50 and the 50-10 ratios, even if in this latter case it is not statistically different from zero. As expected, returns from abstract task increase over time, especially in the upper tail of the distribution. Another very interesting finding concerns the pattern of the service task intensity, which is decreasing along the wage distribution. This means that had only the returns from service tasks changed over time, there would have been an increase in wages in the lower part of the wage distribution, with a polarization effect on the lower tail of the distribution that reduces the 50-10 ratio. This is consistent to what happened in the United States, where the polarization in the lower tail has been basically driven by the service sector (Autor and Dorn, 2012; Mazzolari and Ragusa, 2012). A similar finding is observed for routine, although it is not statistically different from zero. Another interesting pattern is associated to the education variable: had only the returns to education changed over time, wages would have decreased along all the wage distribution, with a peak around the 70th percentile (Figure 10). While been at odds with findings for the US, this is consistent with recent papers that observe falling returns to education in Spain (Izquierdo and Lacuesta, 2006), France (Charnoz et al., 2011), and Italy (Naticchioni et al., 2010), countries that are included in our sample. This is also consistent with the OECD (2011) finding that education entails a reducing impact on inequality trends in the last decades. As far as the other components are concerned, they are almost all very close to zero and/or constant along the wage distribution, apart from the demographic component that displays an inverse U-shape pattern, favoring more individuals who are in the middle of the distribution (Figure 10. In Table 7 we report the decomposition of the three standard inequality indexes (90-10, 9050, 50-10) and the related standard errors, computed bootstrapping the whole procedure (100 replications). [Figure 10 about here.] 18

[Figure 11 about here.] [Table 7 about here.] In the industry level analysis we showed that the technological impact was mainly driven by the service sector, and in particular by the financial and postal-telecommunication industry. To investigate this issue using individual level data, we apply the decomposition analysis to the service sector. Figure 12 shows the detailed wage structure effect only for the service sector. The technological impact is similar to the one detected for the whole economy, with a u-shape patterns. Interestingly, this impact is stronger along the whole wage distribution, and especially in the lower tail of the distribution where the increase of the 10th percentile of the distribution is about 14%, while it was about 8% for the whole economy. This evidence suggests that also using individual data the technological impact seems to play a major role in the service sector. [Figure 12 about here.] Robustness checks As first robustness check we consider the current gross hourly wages in PPP. For sake of space we focus on the detailed wage structure effect, which represents the most interesting finding of the individual level analysis. Figure 13 shows the detailed wage structure effect when using gross current hourly wages in PPP. It emerges that trends are very close to the ones derived using wages non in PPP, with the main difference being that from the median to the top of the distribution the technological impact is constant while before it was slightly increasing. Also the decomposition of the technological tasks provide findings very similar (see Figure 14). [Figure 13 about here.] [Figure 14 about here.] The second robustness check concerns the fact that in Europe there is an important heterogeneity concerning wage dynamics across countries, and that not controlling for this heterogeneity might affect the identification of the relation between wages and task measures. To deal with this issue we introduce in the RIF-regression country dummies as additional covariates, to control for all unobserved differences across countries. Figure 15 shows that introducing countries dummies do not affect the wage structure impact of the technological component, which is even more U-shaped 19

especially in the upper tail. Also the impact of the technological tasks does not change much (see Figure 16. This evidence is in line with the one found in the industry level data analysis. While there is an important heterogeneity in wage trends across countries, the technological impact on wages does not depend much on such heterogeneity. [Figure 15 about here.] [Figure 16 about here.] 3.3.2

Analysis for the decomposition of the dynamics of the gross yearly earnings

Figure 17 shows the decomposition results for the sample of countries for which the gross yearly earnings is available (AT, BE, DK, ES, FI, FR, GR, IE, IT, LU, PT, UK). As already underlined in the descriptives statistics, the gross yearly earnings decreased in the lower tail of the distribution and increased in the upper tail, entailing an increase in earnings inequality. As for the aggregate composition effect, it is much flatter -and always below- than the total change: had only the composition changed over time, wages would have increased much less than what observed. This also suggests that the wage structure effect has to be always greater than the total observed change, counterbalancing the composition effect.15 [Figure 17 about here.] Among the components of the composition effect, Figure 18 shows that most of the patterns are close to those derived for the gross current wages. More specifically, the only positive component is education, which is constant along the wage distribution and anyway quite low in magnitude (at around 0.03), while the other components, apart institutions, are rather constant and very close to zero. Interestingly, the main difference with respect to the gross hourly wages is that institutions is very important in the lower part of the distribution, representing the most important finding related to the composition effect. Had only the share of part time and temporary jobs increased over time, the lower tail of the wage distribution would have shift downward, strongly deteriorating 15

Also in this case both the reweighing and specification errors, computed as defined in the methodological section,

are not statistically different from zero, even if especially in the upper tail of the distribution they amount to a non negligible -0.019.

20

the wages of unskilled workers at the 10th percentile and generating an increasing impact on the 50-10 ratio dynamics, ceteris paribus. This suggest that all the reforms introduced in the last 15 years in Europe aiming at increasing the labour market flexibility have entailed a negative wage impact in the lower part of the distribution. It is not surprising that the impact of institutions emerges when the gross yearly earnings are considered, while no impact is detected when using hourly wages. This result is consistent to the OECD finding in (OECD, 2011): in developed countries the increase in flexibility have generated a positive impact on wage inequality, mainly through a negative impact in the lower tail of the distribution. However, these reforms might have exerted at the same time a negative impact on labour market inequality through a positive (negative) impact on employment (unemployment) rates. It is not easy to recover the prevailing effect, i.e. whether the increase in flexibility has generated an overall positive or negative impact on labour market inequalities. This analysis goes beyond the goal of this paper. [Figure 18 about here.] As for the detailed wage structure effect (19), the patterns are close from a qualitative point of view to those observed for the sample of gross current wages. In particular, the technologic component displays still a U-shape shape, even if at the top of the distribution it slightly decreases. When decomposing the three technological tasks, patterns are similar, with service (and routine) that drives the lower tail of the distribution, and with abstract that displays an increasing trends up to the 80th percentile, and then declines (Figure 20). Consistently with the analysis on current hourly wages, the education component is slightly negative and constant along the distribution, confirming that in Europe education does not entail an increasing impact along the wage distribution.16 [Figure 19 about here.] [Figure 20 about here.]

4

Conclusion

Recent literature shows that wage and job polarization trends are at work for the European countries. In this paper we investigate whether there is wage polarization trends at work in Europe. In 16

We carried out the same robustness checks used for the gross current earnings. Results confirms the main findings

of the baseline specification. For sake of synthesis we do not include them in the text.

21

particular we investigate the trends of unconditional wages as well as the conditional impact of technology on wages, in such a way testing the routinization explanation. We produce both a industry level and individual level analysis. In the industry level analysis, we use the EU KLEMS aggregate industry data, as in Michaels et al. (2010). We point out that there is evidence of the impact of ICT intensity on job polarization, while the evidence on the relation between wage polarization and ICT is much weaker. In the micro evidence we make use of harmonized micro data from two different sources, the ECHP data for 1996 and the EU-SILC data for 2007, harmonizing the definitions of the variables of interest. We make use of two samples of countries. The first one includes Austria, Greece, Ireland, Italy, Portugal, Spain, UK, countries for which it is available the current monthly gross wage. The second one include all Euro15 countries except Netherland and Sweden, for which the gross yearly earnings are available. We also augment this data with task measures that can be considered as a proxy for technological change (Abstract, routine, service) and globalization (offshore). We get these measures from the paper of Goos et al. (2010). By applying the decomposition approach developed by Fortin et al. (2010), we derive the following findings. First, we show that wage inequality increased in Europe, both using gross current hourly wages and gross yearly earnings. However, no wage polarization trends are at work in Europe. Second, we point out that the wage structure related to technological task measures entail a U-shape impact on the changes in the wage distribution, especially in the lower tail of the wage distribution. This U-Shape impact is mainly due to the service task for the lower tail and to the abstract task for the upper tail of the distribution, while the impact of offshoring is constant along the wage distribution. Interesting, education entails a negative impact on changes in inequality, consistently with OECD (2011), and institutions play a role in explaining the drop in the lower tail of the distribution of the gross yearly earnings. The other components (demographic, sector, education) play a less relevant role.

22

References Atkinson, A. (2008): The Changing Distribution of Earnings in OECD Countries, Oxford University Press. Autor, D., L. Katz, and M. Kearney (2006): “The polarization of the US labor market,” The American Economic Review, 96, 189–194. Autor, D., F. Levy, and R. Murnane (2003): “The Skill Content of Recent Technological Change: An empirical exploration,” Quarterly Journal of Economics, 118, 1279–1333. Autor, D. H. and D. Dorn (2012): “The Growth of Low Skill Service Jobs and the Polarization of the U.S. Labor Market,” Working papers, MIT. Barsky, R., J. Bound, K. Charles, and J. Lupton (2002): “Accounting for the black-white wealth gap,” Journal of the American Statistical Association, 97, 663–673. Blinder, A. (1973): “Wage discrimination: reduced form and structural estimates,” Journal of Human resources, 8, 436–455. Brandolini, A. (2007): Measurement of income distribution in supranational entities: The case of the European Union, Banca d’Italia. ¨ Novo (2009): When supply meets demand: Wage inequality in Portugal, Centeno, M. and A. IZA no.4592. Charnoz, P., E. Coudin, and M. Gaini (2011): “Wage inequalities in France 1976-2004: a quantile regression analysis,” INSEE working paper, no.6. Di Nardo, J., N. Fortin, and T. Lemieux (1996): “Labor market institutions and the distribution of wages, 1973-1993: a semi-parametric approach,” Econometrica, 64, 1001–1044. Dustmann, C., J. Ludsteck, and U. Schnberg (2008): “Revisiting the German Wage Structure,” Quarterly Journal of Economics, 142, 843–881. Feenstra, R. and G. Hanson (1999): “The Impact of Outsourcing and High-Technology Capital on Wages: Estimates For The United States, 1979-1990,” Quarterly Journal of Economics, 114, 907–940.

23

Firpo, S., N. Fortin, and T. Lemieux (2011): “Occupational Tasks and Changes in the Wage Structure,” IZA working paper no.5542, revised and resubmitted to American Economic Review. Fortin, N., S. Firpo, and T. Lemieux (2010): “Decomposition methods in economics,” NBER Working Paper, published in Handbook of Labor Economics, Elsevier, 2011. Fortin, N., T. Lemieux, and S. Firpo (2011): “Decomposition methods in economics,” Handbook of Labor Economics, 4, 1–102. Frick, J. and K. Krell (2010): “Measuring Income in Household Panel Surveys for Germany: A Comparison of EU-SILC and SOEP,” SOEP papers on Multidisciplinary Panel Data Research, 265. Goos, M. and A. Manning (2007): “Lousy and lovely jobs: The rising polarization of work in Britain,” The Review of Economics and Statistics, 89, 118–133. Goos, M., A. Manning, and A. Salomons (2009): “Job polarization in Europe,” American Economic Review, 99, 58–63. ——— (2010): “Explaining Job Polarization in Europe: The Roles of Technology and Globalization,” Tech. rep., Katholieke Universiteit Leuven Working Paper, May, revise and resubmit to the American Economic Review. Grossman, G. and E. Rossi-Hansberg (2008): “Trading tasks: a simple theory of offshoring,” American Economic Review, 98, 1978–1997. Hampel, F. (1974): “The influence curve and its role in robust estimation,” Journal of the American Statistical Association, 69, 383–393. Hauser, R. (2008): “Problems of the German Contribution to EU-SILC A research perspective, compar ing EU-SILC, Microcensus and SOEP,” German Council for Social and Economic Data (RatSWD), 80. Izquierdo, M. and A. Lacuesta (2006): “Wage inequality in Spain: Recent developments,” Banco de Espana working paper, no.615. Lemieux, T. (2008): “The changing nature of wage inequality,” Journal of Population Economics, 21, 21–48.

24

Machin, S. (20011): “Changes in UK Wage Inequality Over the Last Forty Years,” in P. Gregg and J. Wadsworth (eds.) The Labour Market in Winter - The State of Working Britain. Mazzolari, F. and G. Ragusa (2012): “Spillovers from high-skill consumption to low-skill labor markets,” fourthcoming in Review of Economics and Statistics. Michaels, G., A. Natraj, and J. Van Reenen (2010): “Has ICT Polarized Skill Demand? Evidence from Eleven Countries over 25 years,” NBER Working Paper. Milanovic, B. (2006): “Global income inequality,” World Economics, 7, 131. Naticchioni, P., A. Ricci, and E. Rustichelli (2008): “Wage inequality, employment structure and skill-biased change in Italy,” Labour, 22, 27–51. ——— (2010): “Far away from a skill-biased change: falling educational wage premia in Italy,” Applied Economics, 42, 3383–3400. Oaxaca, R. (1973): “Male-female wage differentials in urban labor markets,” International Economic Review, 14, 693–709. Oaxaca, R. and M. Ransom (1999): “Identification in detailed wage decompositions,” Review of Economics and Statistics, 81, 154–157. OECD (2011): Divided We Stand: Why Inequality Keeps Rising, OECD Publishing. O’Mahony, M. and M. Timmer (2009): “Output, Input and Productivity Measures at the Industry Level: The EU KLEMS Database*,” The Economic Journal, 119, F374–F403.

25

A

Appendix: Micro data

To study the dynamic of labor earnings in Europe in recent years, we use wave 1996 of ECHP and wave 2007–4 of EU–Silc, as explained in subsection 3.1. Both surveys collect information on current gross monthly earnings on main job, relative to the month in which respondent is interviewed, and on the annual earnings referred to the calendar year previous the interview. Current gross monthly earnings are defined as the monthly amounts earned by employees in the main job, including usual paid overtime, both in ECHP (variable PI211MG) and in EU–Silc (PY200G). From gross monthly earnings we derive hourly wage by multiplying them by 12/52, and dividing by the the number of hours usually worked per week in the current main job (PE005A for ECHP, PL060 for EU–Silc). However, this variable is available in both surveys only for AT, ES, GR, IE, IT, PT and UK. Then we consider annual earnings (variable PI111 for ECHP, PY010G for EU–Silc), which allow to consider all EU15 countries, with the exception of Netherland and Sweden.17 Annual earnings include monthly wages, 13th and 14th salary, extra payments for overtime, holiday pay, earnings from an additional job, other earnings not specified separately, and lump sums payments. EU–Silc provides annual earnings gross of income taxes and social security contribution, while in ECHP amounts are net, with the exception of FI and FR. To compare earnings over time, we need to convert ECHP net amounts to gross, using the net/gross factor (HI020), which has the shortcoming of being constant for all income components. On the contrary, current monthly earnings are always gross, both in ECHP and EU–Silc, even if it is available for a smaller set of countries. Based on these considerations, we decide to use both definitions of earnings in our analysis, in order to enforce our findings and/or to enlighten possible discrepancies when using different definitions of earnings. We also consider a set of covariates of interest, to better interpret changes in earnings inequality in Europe during the period observed. The covariates selected are: gender, education level, type of contract (permanent or fixed–term), whether employee is working full or part–time, industry and potential labor experience. The definitions of these covariates are largely comparable between the two surveys. In few cases, namely industry and potential labor experience, there are some differences. As for 17

Actually, data on annual earnings for NL are available both in ECHP and EU–Silc, but some of the covariates of

interest are missing, so that we exclude NL from our sample.

26

industry sector, in both sources industry classifications, based on NACE–REV 1.1, are more detailed than the one we use. However, they are not fully comparable. For instance, in ECHP (variable PE007B) information about section section D, Manufacturing, is available on its own, while in EU-Silc (variable PL110) section D is grouped with section C and E, Mining and Electricity. Hence we have grouped some sectors in order to match the two classifications. Eventually we obtain the following classification with seven macro-sector: Manufacturing, Wholesale, Restoration and Transport, Financial Intermediation and Business Activities, Public Administration, Education and Health, Others Services. Another relevant covariate is the potential labor experience, considering the age at which one attained the highest level of education. ECHP provides the age when the highest level of education was completed (PT023). If this variable is missing, but the highest level of education attained was less than secondary education (ISCED 0–2, variable PT022), we assume that respondent potentially could began to work at the age of 14. As for EU–Silc, potential labor experience is based on the year when highest level of education was attained (PE030). The flag variable of PE030 identify those that have never been in education. For these respondents we assume that they could not start to work earlier that at the age of 10. Education level is provided by variable PT022 for ECHP, and PE040 for EU–Silc. In both surveys, educational attainments are defined following the International Standard Classification of Education (ISCED). In ECHP the variable has three levels, “lower than upper secondary education”, “upper secondary education completed”, and “first stage of tertiary education completed”, which correspond to ISCED levels 0–1, 2 and 3–5, , respectively. Then we have grouped the more detailed ISCED levels provided by EU–Silc into these three categories. As for the remaining covariates, respondent’s genre is derived from variables PD004 for ECHP and PB150 for EU–Silc; the type of contract, of limited or unlimited duration, is collected from PE024 for ECHP and from PL140 for EU–Silc; variables PE005c for ECHP and PL030 for EU–Silc allows to distinguish between full–time and part–time workers. Finally, we consider some variables that measure the skill requirements at the occupation level, and an index of the offshorability of each occupation. The description of these variables and of the related data sources are provided in subsection 3.1. Note that there are some discrepancies between occupation classifications in Goos et al. (2010) and in ECHP and EU-Silc. All classification refer to the ISCO 2-digit. However, differently from Goos et al. (2010) and from EU-Silc, in ECHP some occupations are grouped together. For instance, ISCO 11 (Legislators, senior officials and managers) and ISCO 12 (Corporate managers) are merged in a single occupation. Hence, for 27

those cases we obtain the corresponding measures computing the weighted mean of the values of Routine, Abstract, Service and Offshorability, with weights given by the share of employees in each occupation, provided by Eurostat.

28

Social

Postal

Health

29

120

100 80

60

40

5 20

0 0

Postal

Public

Auto Sale

Health

Hotels

Construction

Storage

Financial

Postal

Auto Sale

0

Public

ESP

Social

−50

Education

0

Real Estate

0

Construction

120

Health

Financial

25

Postal

UK

Storage

30

Financial

Storage

10

Real Estate

GER

Auto Sale

10 Social

DNK

Public

15

Health

20

Construction

40 Education

20

Real Estate

60

Education

100

Social

Financial

Postal

Real Estate

0

Auto Sale

20

Public

FRA

Social

30 Hotels Social Education Auto Sale Renting Retail Health Real Estate Public Construction Storage Whoolesale Postal Financial

Postal

Financial

Public

Real Estate

Auto Sale

Social

Storage

80

Education

Health Education Construction

140

Construction

Financial

Real Estate

Public

Auto Sale

Storage

Construction

Education

Health

change in ICT capital over value added (pts. pct.)

AUT 250

FIN

150

100 50 200

50 150

40 100

−50 50

0

sector

Figure 1: By-industry 1980-2005 changes of ICT capital compensation over value added. Industries are divided in tradable and non-tradable. Tradeable industries. Source: EUklems.

30 Electrical

GER

1.5 3

1.0 2

0.5 1

0.0 0

Electrical

ESP

Food

0

4

Chemical

1

2.0

Agri

Transport

Recycling

Chemical

Textiles

Food

Machinery

Agri

Electrical

Electrical

Paper

Transport

Recycling

Machinery

DNK

Chemical

Wood Chemical

4

Food

0.6

Agri

0

Chemical

FRA

Electrical

3 Metals

0.0

Rubber

2

Mineral

1.0

Food

3

Food

1.5

Textiles

0.5

Agri

Electrical

Chemical

Food

Agri

2.0

Agri

Electrical

Chemical

Food

Agri

change in ICT capital over value added (pts. pct.)

AUT FIN

UK

1.0

0.8 3

2 2

1 0.4

0.2 1

0.0 0

sector

Figure 2: By-industry 1980-2005 changes of ICT capital compensation over value added. Nontradeable industries. Source: EUklems.

change in ICT capital over value added (pct. pts.)

10

8

6

Sector Tradable Non tradable 4

2

Postal

Financial

Paper

Storage

Transport

Machinery

Public

Whoolesale

Rubber

Chemical

Food

Auto Sale

Wood

Electrical

Metals

Recycling

Hotels

Mineral

Textiles

Real Estate

Construction

Health

Renting

Education

Agri

Retail

Social

0

sector

Figure 3: By-industry 1980-2005 changes of ICT capital compensation over value added. Industries are divided in tradable and non-tradable. Country data are aggregated by weighting by 1980 share of each country’s employment. Source: EUklems.

31

32

0

5

10

15

20

25

Public

0.00

0.02

Mineral Retail Metals

0.06

high skilled

0.04

Storage Hotels Recycling Textiles Whoolesale Paper

Rubber Construction

Agri

Education Social Chemical ElectricalMachinery Real Estate Health Wood Transport Food Auto Sale

Renting

0.08

Postal

0.10

Financial

−10

0

10

20

0.00

0.02

0.04

0.06

0.08

Postal

change in ICT capital over value added

Renting

Education

Health Real Estate Chemical Public

medium skilled

Storage

Mineral Transport Auto Sale Recycling Machinery Rubber Metals Electrical Retail Whoolesale Social Paper

Hotels Wood

Textiles Construction Food

Agri

0.10

Financial

−30

−25

−20

−15

−10

−5 Education

Paper

0.00

Agri

0.02

Textiles Wood Food

0.04

Auto Sale Electrical Storage Transport Construction Hotels Machinery Social

Recycling

Real Whoolesale Estate Health Chemical Rubber Mineral Public

Metals

Retail Renting

0.06

low skilled

0.08

Postal

0.10

Financial

Figure 4: Changes in 2005-1980 wage bill share versus changes in 2005-1980 ICT capital compensation over value added, by skill levels. Data by industries are obtained by summing total ICT capital compensation and total value added over countries. Source: EU KLEMS.

change in wage bill share

AUT

DNK

ESP

40

15

25

Financial

Financial

20

Financial

30

10

15 20

10 5 5

10 0 0 -5 0.00

0.05

0.10

0.15

-0.10 -0.05 0.00

FIN

35

0.10

0.15

0.20

-0.05

0.05

0.10

12

20

Financial

25

0.00

GER

25

30

Financial10 Financial

15 8

20

10

15

6

5

10

∆BSHS

0.05

FRA

4

0

5 0.00

0.05

0.10

0.15

0.20

0.25

0.00

0.01

ITA

0.02

0.03

0.04

0.000 0.005 0.010 0.015 0.020 0.025 0.030

JPN

NLD Financial

Financial

20

30

40

25

30

15 20

20 15

10

10 10

Financial

0

5 5

-0.02

0.00

0.02

0.04

0.06

0.08

-0.05

0.00

UK

0.05

0.10

0.00

0.02

0.04

0.06

0.08

0.10

USA Financial

30

25

25

20

20

Financial

15

15

10 10 5 5 0.00

0.05

0.10

0.15

0.00

0.02

0.04

0.06

0.08

0.10

∆KICT /V A

Figure 5: High skilled Changes in 2005-1980 wage bill share versus changes in 2005-1980 ICT capital compensation over value added, by skill levels. Source: EU KLEMS.

33

AUT

DNK

ESP

35

40

30

30

25

30 25 20

20

20

15

10

15

0

10

5

5

Financial0

Financial

-10 0.00

0.05

0.10

0.15

10

-0.10 -0.05 0.00 0.05 0.10 0.15 0.20

FIN

FRA

25

40

20

30

15

20

10

10

5 0

∆BSM S

0.10

0.15

0.20

-10 0.00

0.01

0.02

0.03

0.04

JPN

40

20

Financial

-5

Financial

0.25

ITA

0.10

0

-10

0.05

0.05

5

Financial-20 0.00

0.00

GER

0

-5

Financial -0.05

0.0000.0050.0100.0150.0200.0250.030 20

NLD

30 10

20 0

Financial

10 0 0

-20 -10

-10

-20

-40

Financial -0.02

0.00

0.02

0.04

0.06

0.08

-0.05

0.00

UK

0.05

0.10

Financial 0.00 0.02 0.04 0.06 0.08 0.10

USA

30 10 20 0

10 0

-10

-10 -20

Financial

Financial

-20 0.00

0.05

0.10

0.15

0.00

0.02

0.04

0.06

0.08

0.10

∆KICT /V A

Figure 6: Medium skilled Changes in 2005-1980 wage bill share versus changes in 2005-1980 ICT capital compensation over value added, by skill levels. Source: EU KLEMS.

34

AUT

DNK

ESP

-10 -10

-10 -15

Financial

Financial

-20

-20

-20 -30 -25

-30

Financial -40

-30 -40

-50 0.00

0.05

0.10

0.15

-0.10 -0.05 0.00 0.05 0.10 0.15 0.20

FIN

-0.05

FRA

-15

0.00

0.05

0.10

GER 5

-10

Financial

-20

0

-20

Financial

-25

Financial

-5

-30 -30

-10 -40

-35

-15

∆BSLS

-40 0.00

0.05

0.10

0.15

0.20

0.25

0.00

0.01

ITA

0.03

0.04

0.0000.0050.0100.0150.0200.0250.030

NLD Financial

Financial

0

0.02

JPN -10

-5

-5 -20

-10

Financial

-10

-15

-15

-30 -20

-20 -25

-40 -25 -0.02

0.00

0.02

0.04

0.06

0.08

-0.05

0.00

UK -10

0.05

0.10

0.00 0.02 0.04 0.06 0.08 0.10

USA Financial

Financial -5

-20 -10 -30 -15 -40

-20

0.00

0.05

0.10

0.15

0.00

0.02

0.04

0.06

0.08

0.10

∆KICT /V A

Figure 7: Low skilled Changes in 2005-1980 wage bill share versus changes in 2005-1980 ICT capital compensation over value added, by skill levels. Source: EU KLEMS.

35

Figure 8: Gross current wage: overall change, aggregate composition and wage structure effects

36

Figure 9: Gross current wage: detailed composition effect

37

Figure 10: Gross current wage: Detailed wage structure effect

38

Figure 11: Gross current wage: Wage structure effect for abstract, routine and service tasks

39

Figure 12: Gross current wages: detailed wage structure effect in the service sector

40

Figure 13: Gross current wages in PPP: detailed wage structure effect

41

Figure 14: Gross current wages in PPP: Wage structure effect for abstract, routine and service tasks

42

Figure 15: Gross current wages: detailed wage structure effect with country dummies

43

Figure 16: Gross current wages: wage structure effect for abstract, routine and service tasks, with country dummies

44

Figure 17: Gross yearly earnings: overall change, aggregate composition and wage structure effects

45

Figure 18: Gross yearly earnings: detailed composition effect

46

Figure 19: Gross yearly earnings: detailed wage structure effect

47

Figure 20: Gross yearly earnings: wage structure effect for abstract, routine and service tasks

48

49

4.55 4.55 8.90 14.84 6.26 5.71 6.74 12.76 13.74 15.56 14.50

∆BSHS

4.73 5.37 7.54 8.52 8.66 9.22 10.20 12.41 18.11 25.95 27.70

BSHS

2.77 17.31 2.91 14.67 13.35 −0.89 16.97 21.26 9.73 13.12 −5.34

∆BSM S

87.57 50.51 76.59 52.60 55.03 66.17 48.05 9.65 50.07 28.45 56.72

BSM S

HS ∆ HH

3.41 4.12 6.68 10.53 5.04 2.76 6.17 9.41 12.08 14.70 9.22

− 7.32 −21.85 −11.81 −29.51 −19.61 − 4.82 −23.70 −34.02 −23.47 −28.67 − 9.16

4.79 2.97 4.28 5.12 5.61 5.52 5.94 8.21 12.93 16.96 20.42

HHS H

∆BSLS

7.70 44.11 15.87 38.89 36.31 24.61 41.75 77.93 31.82 45.59 15.57

BSLS

2.45 18.73 6.67 16.67 16.71 3.59 20.20 23.15 14.35 15.07 −0.72

MS ∆ HH

87.40 44.29 76.59 53.87 50.61 59.81 46.56 7.54 51.56 32.21 60.73

HM S H

39.77 49.48 48.06 30.05 29.66 32.24 52.64 46.81 9.89 29.67 31.74

1980 WHS

− 5.85 −22.85 −13.35 −27.20 −21.75 − 6.35 −26.37 −32.56 −26.43 −29.77 − 8.50

−13.27 − 0.21 4.83 17.61 10.15 23.87 − 6.82 −19.99 17.21 4.26 12.39

1980-2005 LS ∆ HH ∆WHS

7.80 52.74 19.14 41.01 43.78 34.67 47.50 84.26 35.52 50.83 18.84

HLS H

−10.57 8.05 6.50 9.87 7.85 9.98 1.58 −16.10 11.91 5.61 3.61

∆WM S

32.99 28.48 24.45 18.17 16.24 19.90 28.21 33.30 6.72 16.35 21.28

WM S

−19.07 4.88 2.59 2.09 5.37 7.41 2.78 − 8.10 11.03 3.47 − 0.61

∆WLS

26.20 20.52 19.52 17.13 11.92 12.72 23.00 22.24 6.05 16.52 18.19

WLS

0.009 0.014 0.024 0.025 0.015 0.009 0.019 0.011 0.016 0.021 0.029

∆ KVICT A

0.013 0.029 0.012 0.018 0.013 0.018 0.010 0.021 0.014 0.015 0.016

KICT VA

0.05 −0.03 0.07 −0.02 0.03 0.02 0.05 0.06 0.10 0.01 0.03

∆ KNVICT A

0.16 0.17 0.16 0.19 0.24 0.19 0.18 0.27 0.20 0.22 0.24

KN ICT VA

−0.01 0.08 0.07 0.14 0.04 0.03 0.05 0.02 0.04 0.11 0.09

∆ KICT K

0.11 0.16 0.07 0.08 0.04 0.09 0.05 0.07 0.07 0.08 0.06

KICT K

Table 1: The first panel reports by-industry means of labor bill share (BHS , BSM S , and BSLS ); hours worked by skill groups over (HHS , HM S , and HLS ) over total number of hours worked in the economy (H); average hourly real wage (in 2005 USD) by skills (WHS , WM S , and WLS ); ICT (KICT ) and non ICT capital (KN ICT ) over value added (V A); the ration between ICT and total capital (K). The second panel shows level changes of the same variables over the period 1980-2005. Means are weighted by 1980 share of each country’s employment.

ITA DNK NLD UK AUT GER FRA ESP JPN FIN USA

Country

ITA DNK NLD UK AUT GER FRA ESP JPN FIN USA

Country

50 157 0.411 Yes Yes

Observations R-squared Country FE Financial and postal

157 0.522 Yes Yes

29.60*** (4.993)

-58.12** (23.52) 24.81** (11.24) -14.13*** (2.738)

(2) BS-MS

157 0.665 Yes Yes

-31.13*** (4.208)

3.711 (19.21) -15.46 (10.40) 10.44*** (2.493)

(3) BS-LS

28.93*** (5.317)

-22.45 (61.01) 23.20* (12.52) -13.79*** (2.784)

(5) BS-MS

-30.88*** (4.468)

0.720 (57.74) -15.19 (11.62) 10.40*** (2.512)

(6) BS-LS 40.34*** (15.02) -8.471 (5.120) 3.134** (1.356) 48.56** (19.10) -0.782 (4.674) -0.547 (0.737) 6.628 (6.574)

(7) BS-HS

140 140 140 157 0.376 0.518 0.665 0.433 Yes Yes Yes Yes No No No Yes Robust standard errors in parentheses *** p