Master Thesis
Have Information and Communication Technologies
Polarized Wages?
– Measuring the Impact of ICT Investment on Wages –
June 2016
MSc International Economics and Business University of Groningen
Faculty of Economics and Business
ABSTRACT
This thesis tests whether ICTs intensity effects wage inequality as predicted by the routinization hypothesis. According to which, technologies substitute highly routinized tasks of middle-skilled workers. In turn high-middle-skilled workers, complementary to new technologies, benefit. While low-skilled workers remain unaffected, because the main tasks cannot be substituted nor complemented. Using panel data for 9 countries over 25 years, we find no support of the routinization hypothesis, the results show a positive effect of ICT intensity on the high-skilled wage bill share and a negative effect on the other two wage bill shares, implying a positive relationship between ICTs and inequality.
Table of Contents LIST OF ABBREVIATIONS ... I LIST OF FIGURES ... II LIST OF TABLES ... II 1. Introduction ... 1 2. Literature Review ... 3
2.1 The Theoretical Relationship Between ICTs and Income ... 3
2.2 Hypotheses ... 8
2.3 The Empirical Relationship Between ICTs and Income ... 9
3. Methodology ... 12
3.1 Data Selection ... 12
3.2 Baseline Regression Model ... 15
LIST OF ABBREVIATIONS
BEA Bureau of Economic Analysis CES Constant Elasticity of Substitution
CT Communication Technology
GLS Generalized Least Squares
ICT Information and Communication Technology
IT Information Technology
ISCED International Standard Classification of Education
LIST OF FIGURES
Figure 1 – Development of Wage Bill Shares by Skill Group and Country ... 17
Figure 2 – Development of ICT Intensity 1980-2004 ... 18
Figure 3 – Development of the Unemployment Rate by Skill Group ... 30
LIST OF TABLES Table 1 – Characterization of Jobs According to the Nature of Tasks Performed ... 6
Table 2 – Skill Levels According to ISCID Level ... 15
Table 3 – Descriptive Statistics ... 16
Table 4 – Summary Statistics by Country ... 19
Table 5 – Correlation Matrix for all Variables ... 21
Table 6 – Fixed Effects Model ... 26
Table 7 – Random Effects Model ... 26
Table 8 – Decomposing Changes in Wage Bills into Wages and Hours (Random Effects) .... 28
Table 9 – Fixed and Random Effects Mode with Unemployment by Skill Level ... 31
Table 10 – Testing Heterogeneity in Coefficients Across Countries (I) ... 32
Table 11 – Overview of Data Sources... 43
Table 12 – Joint And Individual Significance Tests of the Explanatory Variables ... 44
Table 13 – Multicollinearity Test - Variance Inflation Factors... 44
Table 14 – Wooldridge Test for Autocorrelation in Panel Data ... 44
Table 15 – Testing for Fixed and Random Effects ... 45
Table 16 – Specification Test for Fixed and Random Effects ... 45
Table 17 – Decomposing Changes in Wage Bills into Wages and Hours (Fixed Effects) ... 45
Table 18 – Testing Heterogeneity in Coefficients Across Countries (II) ... 46
1. Introduction
In our “new economy” electronic devices, and the internet have not only found their way into our every day’s life, but also into our workplace. Meanwhile the impact of the automation and computerization on the income are not as easily apparent. Thus, for several years great effort has been devoted to the study of the impact of technological change on the labor market, and specifically on employment structures (job polarization) and wage structures (wage polarization). Since the late 1970s within-country inequality has been increasing in many countries (OECD 2011a). At the same time investments in ICTs have been rising, putting technological change once again in focus of economic research and giving reason to economists to hypothesize that higher investments in ICTs have been a main driver of inequality.
In literature several theories have been proposed that serve as a basis to explain the association between technology and wage inequality. An influential explanation for this positive relationship is the skill-biased technological change hypothesis, according to which the demand for high-skilled workers rises relative to lower educated workers due to technology, changing the occupational employment structure. This so called canonical model offers an explanation for an increased demand of high-skilled workers and rising college wage premium, but rather describes an “uniform shift in employment away from low-skilled and toward high-skilled occupations” (Goos, Manning, and Salomons 2009, 58). Therefore, it cannot explain job polarization and wage polarization. Nevertheless, the notion of skills / human capital was introduced, and still plays an important role in explaining labor market trends.
One prominent hypothesis explaining these recent labor market trends is the routinization
hypothesis, where the degree of routinization in tasks plays a central role in determining the
Technological change has shown its disruptive power on the labor markets in the past, where the hopes of increased productivity and higher wages and the fears of increased inequality and displacement arose at the same time. Technological revolutions induce a “race against the machine”(Tinbergen 1974; Goldin and Katz 2008) for some workers, with the ultimate fear of increased inequality and displacement arose at the same time. Technological revolutions induce a “technological unemployment”, a term introduced in Keynes well-known essay from 1930 “Economic Possibilities for our Grandchildren” (Keynes 2008). Accordingly the “Age of Information and Communication Technologies” can be seen as the fifth technological revolution (Perez 2002). Like in previous technological revolutions marked by telegraphs, steam engines and electric motors, ICTs are also expected to bring about similar increases in productivity (Brynjolfsson and Hitt 2000; OECD 2004). Indeed, current research predominantly concludes, that investment in ICTs is positively and significantly associated with labor productivity growth (Cardona, Kretschmer, and Strobel 2013).
Today, modern innovations and foremost ICTs change the nature of jobs again. With rising wage inequality in todays “Age of Information and Communication Technologies”, the fear of automation is triggered once more. These fears often reflect the fear of displacement and refer to a substitution of labor with technology, which reduce the labor demand for these workers. Hence, technological change remains a sensitive topic for both workers and policy makers. At the same time optimists of digital transformation argue that technology is sometimes also complementary to labor increasing its productivity and demand. As Brynjolfsson & McAfee summarize:
“There’s never been a better time to be a worker with special skills or right education,
because these people can use technology to create and capture value. But there’s never been a worse time to be a worker with only ordinary skills and abilities to offer, because computers, robots and other digital technologies are acquiring these skills and abilities at an
extraordinary rate.” (Brynjolfsson and McAfee 2014, 11)
The before presented routinization hypothesis will function as basis for the hypotheses and the subsequent research question:
Did Information and Communication Technologies (ICTs) attribute to increasing wage inequality between different skill groups?
The remainder of the paper is organized as follows: In chapter two I will first outline the literature on the relationship of technological change and inequality from a theoretical perspective, based on which I will formulate the research hypotheses. Then I will describe the current state of research by giving an overview of the empirical studies. In chapter three I will present the data selection as well as the methodological framework that serves to answer the above defined research question. In chapter four I will contemplate the descriptive statistics and discuss the empirical results, which will undergo a sensitivity analysis. Furthermore, the thesis discusses the main finding and addresses some limitations and suggestions for further research in the concluding chapter five.
2. Literature Review
This chapter summarizes the current literature on the effects of technological change of ICTs by first taking a theoretical perspective, which serves as a foundation for formulation of the hypotheses in 2.2 and the research design in the following third chapter. In order to present the current state of scientific knowledge, I will outline the empirical perspective of the literature, which will act as reference for the discussion of the results later on.
2.1 The Theoretical Relationship Between ICTs and Income
In the neoclassical model the distribution of factor shares depends upon the factor endowments and the underlying technology of the production. The simple Cobb-Douglas Production function describes the relationship between the output Y produced and the inputs X needed in its production:
𝐹 [𝐾, 𝐿, 𝐴] = AKα L1-α 𝑤𝑖𝑡ℎ 0 < 𝛼 < 1 [1]
K represents the total capital invested in production and L the labor. The elasticities of output
of the factors K and L are defined by α and 1-α respectively. The factor A is a measure of the Total Factor Productivity, which often represents the technological progress. It is therefore
to output.1 The elasticity of substitution becomes crucial, because when it imperfect (< 1 against
neoclassical assumption), technological change is either capital augmenting or labor augmenting. For many this implies a capital versus labor relation, usually assessed by looking at capital and labor shares of national income and reflecting the fear of automatization (Acemoglu 2009, 656). As Acemoglu declares
“[t]he study of why technological change is sometimes biased towards certain factors or sectors is both important for understanding the nature of endogenous technology and also because it clarifies the distributional effects of technological change” (Acemoglu 2009, 497).
This mainly refers to skill-biased technological change, which “is frequently cited as the leading cause of growing wage inequality since the 1980’s”, as noted by Zoghi and Pabilonia (2007, 1).
Canonical Model
Again, using a simple demand and supply model, we can introduce human capital H, the stock of skills embedded in labor. Taking into account the heterogeneity of labor, the adapted production function changes to:
Y(𝑡) = 𝐹 (𝐾(𝑡), 𝐻(𝑡), 𝐴𝐿(𝑡) [2]
This allows to investigate the proximate sources of income differentials between heterogeneous labor, namely physical capital, human capital and technology (Acemoglu 2009, 118).
In factor-biased technological change, technological progress leads to increases in the marginal productivity of one factor relatively to the other factors. Accordingly, the skill-biased technology hypothesis refers to a relative increase in the marginal productivity of high-skilled labor, resulting in a relative increase of high-skilled worker’s wages. Thus, technological change has an influence on the relative demand of labor as well as on its relative price, the wage premium. Labor-biased technological change increases the relative demand for high-skilled labor and increases the relative marginal product of labor: 2
Marginal Product of Labor = 𝛼 𝑌
𝐿 [3]
11 Labor augmenting technological progress refers to “an increase in technology A(t) increases output as if the economy had
more labor” and correspondingly capital augmenting, refers to an increase in output “equivalent to the economy having more capital” which increases the productivity of the factor (Acemoglu 2009, 83).
2 In an equilibrium setting of supply and demand for labor, the compensation of labor should be according to the marginal
ICTs are then complementary for high skilled workers and substitute for low skilled workers, which leads to the fact that the labor shares of those skill groups change accordingly. The Constant Elasticity of Substitution framework (CES), which was inspired by (Tinbergen 1974) linking the relative demand for skills to technology, is usually taken as a theoretical starting point for the econometric analysis. In this canonical two-factor model the two inputs of high-skilled workers (AH) and low-skilled workers (AL) determine the aggregate output:
𝑌 = [ ( 𝐴𝐿 𝐿)𝜎−1𝜎 + (𝐴𝐻 𝐻) 𝜎−1 𝜎 ] 𝜎 𝜎−1 ⁄ [4]
When the aggregate elasticity of substitution σ is bigger than one (σ > 1), then high and low-skilled workers are gross substitutes and a relative increase of AH to AL raises the relative
marginal product of H, thus the wages for high-skilled workers. Reversely, when σ < 1 they are complements and wages decrease if the supply of high skilled workers rises. Skill-biased technological change increases AH / AL, and the skill-premium rises and with this inequality
increases as well. The CES approach “provides an interpretable structural framework to analyze between- and within-industry demand shifts for multiple skill groups” (Katz and Autor 1999, 1517). The critique of the canonical model is based on its inability to explain United States labor market trends of the 1990s and 2000s; on one hand the job polarization as a result of increasing employment levels of high- and low-paid occupations and decreasing employment of the middle and on the other hand the rise inequality, as mirrored in U-shaped wage-growth in skill-percentiles (Autor and Dorn 2013, 1554).
The Routinization Hypothesis
TABLE 1 – CHARACTERIZATION OF JOBS ACCORDING TO THE NATURE OF TASKS PERFORMED
Acemoglu and Autor (2011) link the occupational changes to the characterization of the job tasks as presented in table 1. These tasks characteristics are then linked to the different skill levels by looking at the task intensity performed and the educational level of the workers. Middle-skilled workers, such as Bookkeepers and Clerks, perform many tasks that are routinized and require cognitive and analytical thinking (I). Due to their nature they can more easily be substituted by new technology. At the same time jobs that are intensive in highly routinized and manual tasks (II) can also easily be automated. These jobs are also performed by middle-skilled workers (e.g. machine operators). Increased investments in ICT thus decrease the labor market price of routine-tasks and consequently the middle-skilled workers with a comparative advantage in routine tasks will see their wages fall (Acemoglu and Autor 2011, 1153-1154).
The work of high-skilled labor (e.g. of teachers and managers) highly involve abstract, non-routine tasks (IV), where ICTs are complementary. Thus, high-skilled workers benefit from the use of ICT and gain in comparative advantage, because information is made available to them at a lower cost and they can more effectively use their analytical skills. In effect high-skilled workers increase their productivity (Acemoglu and Autor 2011).
In contrast to that, for low-skilled labor increases in productivity cannot be expected, because there are limited opportunities to complement their core tasks. The manual tasks (III) performed by the low-skilled labor require face-to-face contact and therefore are not directly affected by technological change.
The wages of high-skilled workers are expected to increase due to the gain in comparative advantage over the medium skilled workers. The medium-skilled workers’ comparative advantage for previously performed tasks erodes, so that they will perform fewer tasks and eventually lead to a total displacement of these workers. In this line of thought, unemployment
Source: Own presentation, based on Acemoglu, Autor (2011)
is a “side effect of progress”, as called by Baumol and Wolf (1998), who link technological change to an increase in the level and the duration of unemployment in their analysis. Hence, unemployment is linked to labor-saving technological change, when labor is displaced by ICTs. Acemoglu (1999) provides a framework to analyze the effects of a changing composition of the workforce on wages and unemployment. By applying this framework to the United States, he concludes that technological change has led to increasing unemployment rates for all educational groups. As indicated by the Krugman hypothesis the decline in the demand for unskilled workers has “two sides of the same coin” (Krugman 1994, 37): a trade-off between rising wage inequality or rising unemployment, where the latter was identified to have occurred in continental Europe and the other is specific to Anglo-Saxon countries. Key to this difference is the rigidity of the labor market, where the causal argumentation is that inflexible workers are subject to a higher unemployment risk.3 Thus, in addition to technological change in form of ICTs, the changing role of labor market institutions is an important factor in explaining the increasing inequality.
Western and Beckett (1999) investigate differences between the European and the US Labor Market, in order to offer an explanation for the comparatively low unemployment rate in the 1980s and 1990s. They challenge the view, that the deregulation of the US labor market and the low unemployment are linked and argue that changes in the penal system caused an underestimation of the unemployment rate in the long run. Furthermore, higher rates of imprisonment raise unemployment in the long run and therefore lead to increasing inequality. Western and Beckett’s analysis thus does not only link unemployment and inequality, but also points out that firstly United States unemployment measures hide the joblessness in the United States and secondly, that the United States is not unregulated since incarceration can be seen as a labor market intervention, thereby they stress the central role of institutions in shaping the labor market.
From a theoretical point of view unions, effecting (minimum) wages through collective bargaining and by influencing political decision making, can have a compressing effect on wage differentials as the analysis in “Unionism and the Dispersion of Wages” by Freeman (1980) suggests. Card (1996), using longitudinal data, identifies a relatively stronger union wage effect for the lower educated groups than for higher educated groups. In a more recent study Card, Lemieux, and Riddell, looking at micro-data and take into account the heterogeneous skill
levels of workers, conclude that “unions have an equalizing effect on the dispersion of wages across skill groups” 4 (2004, 555), supporting the Freemans key findings.
2.2 Hypotheses
In chapter 2.1 the link between ICT and wage inequality was portrayed from a theoretical perspective, with particular attention paid to the routinization hypothesis as a prominent explanation for this positive association. Based on the theoretical implications of the routinization hypothesis, individual hypothesis for the impact of ICT investments on each skill group are deducted.
Hypothesis 1: There is a positive association between ICT investments and wages of
high-skilled workers.
Both the hypothesis of skill-biased technological change and the routinization hypothesis suggest that due to technological change induced by higher ICT investments, the relative wages of high-skilled workers will increase, because they are complementary to ICTs and therefore gain comparative advantage over other, lower-skilled workers.
Hypothesis 2: There is a negative association between ICT investments and wages of
middle-skilled workers.
Wages of middle-skilled workers decrease, because the routine tasks they perform can be substituted by new technologies. Accordingly, middle-skilled workers with a comparative advantage in routine-tasks, both cognitive and manual, will see their wages decline, because the labor market price for routine tasks will fall due to increasing investment in ICTs.
Hypothesis 3: There is no association between ICT investments and wages of low-skilled
workers.
Wages of low-skilled workers are largely unaffected by ICT investments, since (so far) ICTs cannot substitute nor complement the main non-routinized tasks of low-skilled workers.
4 Even though, one hast to note, that there is a strong difference between the impact of unionization on wages for men and
The above defined hypotheses rely on the findings of Michaels, Natraj, and van Reenen (2014, 63), who in the first part of their most recent paper link task measures of occupations (routine cognitive, routine manual, non-routine manual, and non-routine manual) from Autor, Levy, and Murnane (2003) to the skill levels as defined in the EU KLEMS database.5 They support the assumptions of Autor, Levy and Murnane (2003) that occupations with high shares of high- and low-skilled workers have low intensities both routine-cognitive and routine-manual tasks. Further high-skilled occupations show a high intensity for non-routine cognitive tasks and a low score for non-routine manual tasks, the middle-skilled occupations’ scores are as expected reversed, with high scores for non-routine manual tasks and low scores for non-routine cognitive tasks. As suggested by theory, the scores for low-skilled occupations are more mixed.
2.3 The Empirical Relationship Between ICTs and Income
Technological progress and its impact on the labor process through automation and computerization has been focus of research for many years. With “Labor and Monopoly Capital”, where he investigated the division of labor in the United States, Harry Braverman again sparked the debate in 1974 about automation and skills. As a key element of technological change, the substitution of labor by capital causes a degradation of workers who lose the control over their work. Even though many researchers simply focus on the “deskilling of the labor force” in Braverman’s extensive analysis, but in fact he states “a structure is given to all labor processes that at its extremes polarizes those whose time is infinitively valuable and those whose time is worth almost nothing” (Braverman 1998, 58).
Research in general has documented an increase in inequality since the 1980s within many countries, while a great part of research has been focused on United States inequality and the identification of the drivers behind this alarming tendency. The OECD (2011a) and the IMF (2007) identified policies and institutions, economic globalization, and technological change as possible drivers for inequality. Furthermore, they concluded that technological advances account for a great share of changes in wage inequality (OECD 2011a, 26).
The skill biased-technological change hypothesis has been identified as an explanation for job polarization in the United States until 1980. Empirical research greatly documented increasing wage inequality in the United States and identified technology as a key driver for inequality by
5 Autor, Levy, and Murnane (2003) constructed task scores measuring skills linked to the characterization of the tasks. Being
focusing on changes in the demand for skills due to technological change. The results for other countries are more mixed and only few study the impact of increasing ICT intensity effects on wage differentials between skill groups.
In line with the canonical model these studies examine the skill-biased technological change on the national level, mainly focusing on the US, in order to investigate whether rising use of technologies is a driver for the observed increasing personal income inequality (Levy and Murnane 1992; Bound and Johnson 1992). Juhn, Murphy, and Pierce (1993) and Katz and Murphy (1992), using the same Current Population Survey data, find that other potential explanations, including decline in manufacturing, in unionization and in growth in college-educated population cannot explain the variations in wage structures on national level over time, and therefore conclude that “the unobserved variable driving the inequality must be technological change” (Brown and Campbell 2002, 8).
Early studies, based on the canonical model, only provide evidence for an “college-versus-high-school earnings gap” (Acemoglu and Autor 2012, 13) by showing that skill-biased technological change causes a higher demand for skills which increases the college premium. But they cannot explain changes in the labor market structure, especially the job polarization taking place at the high and low end of the skill distribution, in different time periods (Acemoglu and Autor 2011, 1115). Lindley and Machin also contribute the rise in United Kingdom’s wage inequality to changes in the education structure, namely the rising supply of highly skilled postgraduate workers. Technological change, measured by computer usage and investment is strongly correlated with these relative demand shifts in the Lindley and Machin’s study of the UK.
Levy, and Murane (2003). However, these studies only provide evidence for the routinization hypothesis on job polarization and solely show that a wage polarization coincides with changes in the skill distribution, but do not formally investigate the association.
Inspired by Autor, Levy, and Murnane (2003) and Acemoglu and Autor’s preliminary study of the link between the tasks content and trends in wage distributions, Firpo, Fortin, and Lemieux (2011) investigate the impact of changing task prices on the United States wage structure. In accordance with the routinization hypothesis they predict that occupations with increasing / decreasing market values will see increases and decreases in their remuneration. Linking the task measures to technological change, their findings show a displacement of routine task intensive jobs and furthermore attribute part of the US wage polarization to this fact. Machin (2011) also links United Kingdom’s wage inequality over the last decades to Autor, Levy, and Murnane’s hypothesis of routinization.
Only few researchers target other countries and as Freeman and Katz conclude (1995), increases in inequality in Continental Europe were relatively low in the last three decades, compared to the Anglo-Saxon countries, but more recent research identifies significant increases in European Inequality (OECD 2011a).
Country-studies focusing on Continental Europe also identify an impact of technological change on the labor market. Dustmann, Ludsteck, and Schönberg (2009) confirm wage inequality was increasing in the top half during the 1980s and started to increase in the bottom half from the early 1990s in (West) Germany. Building on the work of Spitz-Oener (2006), they identify that the change in the returns to skills is a result of a shiftz in the relative demand and supply in the skill groups. They conclude that these shifts are driven by technological change. On the one hand, there is further evidence on supply and demand forces causing wage polarization are provided for Portugal (Centeno and Novo 2014) and Spain (Carrasco, Jimeno, and Ortega 2015). On the other, some studies show no evidence for skill-biased technological change and wage polarization for Spain (Lacuesta and Izquierdo 2012), France (Charnoz, Coudin, and Gaini 2011), and Italy (Naticchioni and Ricci, 2011), where findings show wage inequality and returns to skills are decreasing.
in ICT intensity (ICT compensation/value added) is associated with a rise in the wage bill share of high-skilled workers and a decline for the share of middle-skilled workers at the industry level, thus showing a conditional impact of technology in support of the wage polarization hypothesis. Naticchioni and Ragusa (2014), also making use of the EU KLEMS data, investigate the unconditional impact of technology on wages. While they find no evidence for a polarization of unconditional wages across Europe (except for Germany), they conclude the evidence for a conditional impact on wages is more mixed, with a small effect on high-skilled labor and no significant effect on middle and low-skilled labor.
3. Methodology
This chapter sets the outline to investigate the impact of ICT on the wage structure in several European countries, Japan, and the United States as an indication for wage inequality. First, the data selection will be presented and critically assessed. After that, the baseline regression model, that serves as a basis for the empirical research will be discussed.
3.1 Data Selection
The main data source for this paper is the EU KLEMS database, which provides industry- and country-level data on capital (K), labor (L), energy (E), material (M), and service inputs (s). Even though the database is centered on a growth accounting framework, the focus of interest here are the different types of capital inputs and labor inputs.6 The EU KLEMS database
includes measures for 25 EU member states, the United States, and Japan and cover the period from 1970 to 2005. The database uses national statistical institutes and harmonizes the national data according to defined procedures, therefore adjustments for several variables where made. An overview of the data sources is provided in table 11in the appendix.
Due to the data availability the balanced panel includes data from 1980 to 2004 for nine countries, being Austria, Denmark, Spain, Finland, Italy, Japan, Netherlands, the United Kingdom, and the United States of America.
Income distribution Data (Wage Structure)
6 Timmer et al. (2007) and O´Mahony and Timmer (2009) provide a summary an overview of the methodology and construction
Taking the heterogeneity of labor into consideration, the EU KLEMS database offers data on the wage rates by individual characteristics of labor such as age, gender and educational attainment. The thesis focusses on the hourly employee labor compensation by educational attainment. Educational attainment, defined as the highest degree of education an individual has completed, is used as a measure of skill, according to which the worker groups low-skilled (at most a primary education), medium-skilled (at least secondary education), and high-skilled (at least tertiary education) are formed (Timmer et al. 2007, 27). It must be critically noted that this split might be too restrictive, since differences in educational systems across Europe exist (Timmer et al. 2010, 64). In addition to that, the labor compensation of self-employed is not taken into consideration in National Accounts, upon which the EU KLEMS input data is constructed (O'Mahony and Timmer 2009, 380). Using the total compensation and the total hours performed by each skill group, the wage bill shares are computed for each skill group. Wage bill shares were also used in previous studies, such as Michaels, Natraj, and van Reenen (2014), Naticchioni and Ragusa (2014), offering the possibility for comparison.
Technological Change Data (ICT Investment)
For this study the broad definition by the World Bank will be used, which defines ICTs as “goods that are intended to fulfill the function of information processing and communication by electronic means, including transmission and display” (World Bank 2014). Nevertheless, it is important to stress that there is no consensus in the literature upon the definition of ICTs; definitions vary depending on the context and the application of the term (Zuppo 2012). As many others, the OECD defines ICTs in respect to the measurement level and the mechanism through which ICTs are hypothesized to have an economic impact. Thus, the OECD distinguishes between ICT capital, the ICT producing sector, and ICT use (Brynjolfsson and Hitt 2000; OECD 2004). Accordingly, ICTs can be seen as production output of the ICT producing sector, or as a production input, when ICT capital is regarded.
The EU KLEMS database provides data for several asset types, ICT assets include assets in office and computing equipment (IT), communication equipment (CT), and software.7 Stocks for each individual asset are estimated using the perpetual inventory method, where past investments are subject to geometric depreciation rates and weighted according to the relative efficiency of the capital good. The capital stock A for asset type k at time T is given by:
𝐴𝑘,𝑇 = ∑(1 − 𝛿𝑘)𝑡−1 𝐼
𝑘,𝑇−𝑡 = 𝐴𝑘,𝑇−1(1 − 𝛿𝑘) + 𝐼𝑘,𝑇 ∞
𝑡=0
Where the investments 𝐼𝑘,𝑇−𝑡 of each asset type are subject to a constant depreciation rate 𝛿𝑘. One set of depreciation rates from the United States’ Bureau of Economic Analysis (BEA) is used for all countries in a harmonized approach, so that depreciation rates are identified for each asset type and industry. Thus, by assumption all national depreciation rates equal the rate of the United States, which can by criticized as an unrealistic simplification (Timmer et al. 2007, 39). Nevertheless, this approach seems appropriate since further information on depreciation rates is limited.
The ISIC Rev.3 from the November 2009 Release of the EU KLEMS database provides the data for the capital inputs. In this thesis the nominal gross fixed capital formation of ICT assets as a ratio of the nominal gross fixed capital formation of all assets is used as an indicator for ICT intensity, following OECD (2011b). The correct measurement of ICT investment is also facing measurement issues: most importantly, ICT investment does only consider investment in ICT products and does not include ICT embodied in equipment and therefore measurement bias might occur due to differences in the treatment of investment in intermediate products. Especially the rental and licensing of intangibles, such as software, pose a measurement issue on ICT investment (OECD 2011b, 83).
Employment Data (Employment Structure)
As depicted in 2.1, in the simple supply and demand framework technological change effects the demand for skilled labor and positively effects the remuneration of labor with complementariness to technology in theory, when their marginal productivity increases. The changes in the employment structure are accounted for by using the hours worked by persons of a skill group engaged. The KLEMS database again provides the employment data by educational attainment used for the employment by skill group. An advantage of using the hours performed by each skill group is that part-time workers are also considered.8
Unemployment data
As explained in 2.1, technological change and unemployment are theoretically linked when technological change is labor replacing. Therefore, the unemployment rate is used to control
for this impact. The data is provided by the OECD database, where the unemployment rate is defined as the number of people unemployed as a percentage of the civilian labor force (armed forces excluded). Additionally, for our sensitivity analysis unemployment rate by educational attainment is used from the Eurostat database, reflecting the unemployed persons as a percentage of the active population by educational attainment, as defined by International Standard Classification of Education (ISCED) in 2011. Like in the EU KLEMS database the educational attainment will be used as a measure for the skill level as depicted in table 2.
Unionization
Unions can affect wages through the power of collective bargaining in wage negotiations and they can assert their power on politics and push minimum wages. This effect is controlled for by including the union density rate reflecting the ratio of wage and salary earners that are trade union members, divided by the total number of wage and salary earners. The data is taken from the OECD Labour Force Statistics, which is based on survey and administrative data and is adjusted for non-active and self-employed members.
3.2 Baseline Regression Model
To study the impact of ICT investment on the wage bill shares of low-, middle- and high-skilled groups and controlling for the unemployment and unionization density, the following baseline panel regression models will be used:
𝐵𝑆𝐻 𝑖𝑡 = 𝛼 + 𝛽1 𝐼𝐼𝐶𝑇 𝐼𝑇𝑜𝑡 𝑖𝑡 + 𝛽2 𝑈𝑛𝑒𝑚𝑝𝑖𝑡 + 𝛽3 𝑈𝑛𝑖𝑜𝑛𝑖𝑡+ 𝜀𝑖𝑡 [5] 𝐵𝑆𝑀 𝑖𝑡 = 𝛼 + 𝛽1 𝐼𝐼𝐶𝑇 𝐼𝑇𝑜𝑡 𝑖𝑡 + 𝛽2 𝑈𝑛𝑒𝑚𝑝𝑖𝑡 + 𝛽3 𝑈𝑛𝑖𝑜𝑛𝑖𝑡+ 𝜀𝑖𝑡 [6] 𝐵𝑆𝐿 𝑖𝑡 = 𝛼 + 𝛽1 𝐼𝐼𝐶𝑇 𝐼𝑇𝑜𝑡 𝑖𝑡 + 𝛽2 𝑈𝑛𝑒𝑚𝑝𝑖𝑡+ 𝛽3 𝑈𝑛𝑖𝑜𝑛𝑖𝑡+ 𝜀𝑖𝑡 [7] TABLE 2 – SKILL LEVELS ACCORDING TO ISCID LEVEL
Skill-level Description ISCID level
Low less than primary
primary
lower secondary education
0 1 2
Middle upper secondary
post-secondary, non-tertiary education
3 4
Where 𝐵𝑆𝑆 𝑖𝑡 = (
𝑊 𝑆∗ 𝑁𝑆
𝑊𝐻 ∗ 𝑁𝐻+ 𝑊𝑀 ∗ 𝑁𝑀+ 𝑊𝐿 ∗ 𝑁𝐿) is the wage share of skill group S = {H,M,L} and
𝑁𝑆 is the number of hours worked by skill group S. In order to test the hypotheses, the baseline
panel regression model [5-7] are used, where the dependent variable measures the wage bill share of each skill group of country i in year t. Wage bill shares are a useful measure, because “each qualification is weighted by its price (the wage)” (Michaels, Natraj, and van Reenen 2014, 63). Hence, using wage bill shares does not only allow for cross-country comparisons, even when qualifications measures between countries differ, but results can also be contrasted to the previous findings by Michaels, Natraj, and van Reenen (2014) and Naticchioni and Ragusa (2014).
The independent explanatory variable of central interest ICT intensity, 𝐼𝐼𝐶𝑇
𝐼𝑇𝑜𝑡𝑖𝑡is the total sum of
investments in ICT assets divided by the investment in total assets in country i in year t, reflecting the technological change towards ICTs. To account for changes in the composition of the employment structure, 𝑁 for each skill group is included. Additionally, controls are
Unempit, the unemployment rate and Unionit, the union density rate.
3.3 Descriptive Statistics
A summary of the statistics for the averaged levels of the key variables from 1980 to 2004 is provided by table three Since we have a balanced panel for 9 countries (groups) over 25 years the number of observation (N) is always 225.
The descriptive statistics already indicate that on average the middle-skilled wage bill share is the highest and the high-skilled wage bill share the lowest across all countries. Looking at the standard deviation we can see an appropriate variation of data for all variables.
TABLE 3 – DESCRIPTIVE STATISTICS Observations Groups Mean Standard
Deviation Minimum Maximum High-Skilled
Wage Bill Share 225 9 10.04 11.19 .36 44.42
Middle-Skilled
Wage Bill Share 225 9 70.88 27.21 1.28 99.31
Low-Skilled
Wage Bill Share 225 9 19.08 25.37 .004 96.68
ICT Intensity 225 9 11.53 4.04 4.76 23.50
Union Density Rate 225 9 39.39 21.87 11.26 80.65
When we look at the cross-country trends of the wage bill share by skill separately in figure 1 below, we can identify that for most countries the bill shares of the middle-skilled is above the shares other two skill groups in the regarded time period of 1980 to 2004. Spain and Finland thereby present an exception: in Spain the wage bill share of the low-skilled is the highest over the whole period and in Finland we can see that the low-skilled wage bill share drops from a relatively high level to a share well below the other two skill groups. Nevertheless, because educational levels indicating skill levels might differ across countries, interpretations of the levels should be regarded carefully. The plots of the wage bill shares of Italy and Japan directly reflect the dominance of the middle-skilled wage bill share. This can be regarded in more detail when we look at table 4.
Across all countries the share of investment in ICT assets as a share of investments in total assets is rising as indicated by figue 2 below. The Anglo-Saxon countries share is clearly above the average country share, eventhough it seems to have declined in the early 2000s. Table 3
will provide further details on the countries means (table 4A) and the relative changes (table 4B) over time of ICT investment and all other key variables.
Table 4A shows the mean levels of the variables for each of the observed countries and table 4B shows the changes from 1980 to 2004 accordingly. Where columns 1, 4, and 9 in table 4 show the dependent variables, the wage bill shares of the three skill groups, column 10 indicates the mean levels of the main independent variables of interest ICT investment over total investment and columns 11 and 12 show the mean levels per country for the control variables union density rate and unemployment rate. The wage bill shares are split into their components the wages and the hours worked, relative terms of total hours are used for a better comparison.
19
TABLE 4 – SUMMARY STATISTICS BY COUNTRY
A. Mean levels by Country
High-Skilled Middle-Skilled Low-Skilled
(1) (2) (3) (4) (5) (6) (7) (8) (9) (10) (11) (12) Country BS 𝑊 𝐻 𝑊𝑇𝑜𝑡 ⁄ 𝐻𝐻⁄𝐻𝑇𝑜𝑡 BS 𝑊 𝑀 𝑊𝑇𝑜𝑡 ⁄ 𝐻𝑀⁄𝐻𝑇𝑜𝑡𝑎𝑙 BS 𝑊 𝐿 𝑊𝑇𝑜𝑡 ⁄ 𝐻𝐿⁄𝐻𝑇𝑜𝑡 𝐼 𝐼𝐶𝑇 𝐼𝑇𝑜𝑡 Union Unemp Austria 2.83 14.56 8.96 84.34 65.35 62.36 12.83 28.67 20.09 7.88 45.1 3.7 Denmark 0.95 8.3 5.16 74.04 62.4 56.62 25.01 38.22 29.3 15.0 75.6 7.2 Spain 9.04 24.51 14.08 11.03 20.65 18.92 79.93 67 54.84 09.5 14.8 17.4 Finland 31.38 37.93 27.61 39.96 34.37 39.77 28.66 32.62 27.7 10.6 73.9 8.6 Italy 1.00 9.5 7.53 98.84 88.2 88.83 0.15 3.65 2.3 9.0 39.4 10.5 Japan 12.17 28.36 18.46 78.6 55.86 61.3 9.23 20.24 15.78 10.1 25.2 3.2 Netherlands 4.54 30.5 7.5 93.57 60.99 82.06 1.88 10.44 8.51 10.9 25.8 7.1 United Kingdom 5.00 19.74 11.19 83.65 62.33 64.36 11.35 24.45 17.93 14.7 39.1 8.1 United States 23.43 38.57 26.15 73.87 52.9 61.04 2.70 12.81 8.53 16.18 15.67 6.26 Mean 10.04 23,55 14,07 70,88 55,9 59,47 19,08 26,45 20,55 11,53 39,39 8,0
B. Changes from 1980 to 2004 by Country
On average the high-skilled bill share is the lowest with 10%, the middle-skilled have the highest share of wage bills on average across countries over the period of 71% and the low-skilled wage bill is around 20%. Table 4B indicates that both the high and the middle wage bill have been increasing by over 12%, accordingly the low-wage bill share has been decreasing over the period. The wage share and the labor shares (in hours) by each skill group offer more details on the composition of the wage bill shares. The high skilled groups receive 24 % of the wages on average while the labor performs 14 % of the total hours. The increases in the wage shares have exceeded the increases in the employment shares in all countries. In contrast to that, in the middle skilled group, which reflects the biggest working group, the share in hours performed increases more than the wage share. The hours performed by lowest skill group is 21% and their wage share is 26 % on average, over time the wage bill share has been decreasing, which is due to decreasing shares in both hours and employment with about the same percentage decrease in the shares.
The descriptive statistics for each country show variation in the wage bill shares reflecting the heterogeneity. Looking at the middle-skilled wage bill share, we can directly identify more cross-country differences; in all countries except Italy, the Netherlands and the US the share has been rising strongly, but in those three countries the middle-skilled wage bill share dropped from 1980 to 2004, which can already be seen in figure 1 above. Whereas in Italy both wage and hour shares are relatively high but decreasing over time, in the Netherlands the hour share, which is already above the decreasing wage share is increasing. Explicitly strong high-skilled wage bill shares can be found in the United States and Finland, where in both cases the mean wage shares are much higher than the employment shares by hours. Over the 25 regarded years the percentage of ICT in total investment (table 4, column 10) has been rising in all countries, the highest increase took place in United Kingdom where the mean is already at a high level. In comparison Italy’s mean level is low and has not been increasing very much over the 25-year period.
share and a stronger negative correlation with the middle-skilled wage bill share and in contrast to that a positive correlation with the wage bill share of the low-skilled.
3.3 Model Specification
With the above defined baseline regression a panel data analysis is performed, because it allows us to test a group of cross-sectional units (countries) over time, accounting for heterogeneity of unobserved individual characteristics (cross-sectional dimension) and intertemporal dynamics (time-series dimension) (Hsaio 2007). Additionally, panel data promises higher efficiency, because more information can be used to produce more reliable parameter estimates than time-series analysis alone. Furthermore, the dynamics of adjustments can be assessed like the adjustment of the labor market to new technologies in this case. But nevertheless “the economist using panel data has to its limitations”, as Baltagi (2008, 11) indicates. A disadvantage of panel data lies in the data collection. Due to the extensive data needed for panel data analysis, other problems, such as measurement errors and selectivity problems, might arise.
This thesis makes use of the EU KLEMS database, providing extensive data on input parameters including investment in assets by types, hours worked and the labor compensation for the period of 1970 to 2004. Even though the database includes 25 European countries, Japan, and the United States, 9 countries remain for the period of 1980 to 2004 due to the availability of data. As indicated before, the skill levels used in the EU KLEMS database do not rely on a homogenous definition and especially among the middle- and low-skilled groups different definitions of the educational levels can lead to biases. Another source for bias could stem from the unclear definition of ICTs, as discussed in 3.1. As mentioned by Baltagi (2008, 9), selection bias selectivity problems are another disadvantage in data that cannot be solved by applying panel data analysis. On one hand selection bias arises from self-selection of the workers in the Labour Force Surveys of each country and on the other hand the selection of data by analysts and data processors induces a selection bias (Heckman 1979). Furthermore, simultaneity may arise, when the independent variables are simultaneously determined by the dependent variable
TABLE 5 – CORRELATION MATRIX FOR ALL VARIABLES
hs_bs ms_bs ls_bs ictinvest union unemp
In order to address the potential endogeneity problems a common strategy is to use lagged explanatory variables as instruments for endogenous explanatory variables. However, as Bellemare, Masaki, and Pepinsky note “without careful arguments on substantive grounds, lagged explanatory variables should never be used for identification purposes” (Bellemare, Masaki, and Pepinsky 2015, 35). Furthermore, their results characterize that even under favorable conditions lag identification not only hides the endogeneity problems but also leads to faulty statistical inferences. Thus, no lags were included in the regressions.
The baseline model was tested for (1) multicollinearity, (2) heteroscedasticity (3) autocorrelation and last but not least the (4) normality of residuals was assessed.
When explanatory variables are perfectly correlated, multicollinearity is present leading to a violation of the least squares assumptions, then the least squares estimator is not the best linear unbiased estimator. The summary statistics did not show any unreasonable variation in the data and the pairwise correlations (table 5) are low giving no indication for multicollinearity. The key explanatory variable is positively correlated with the high- and middle-skilled wage bill shares and negatively with the bill share of the lowest educated. Among the independent explanatory variables (ictinvest, union, and unemp), the correlations are all rather low. Looking at the correlations, we can assume that we can estimate independent effects of each of these variables on the wage bill share, but because we have more than two predictors these results can be misleading. Table 12 in the appendix, displays the joint and individual tests of significance for the independent variables. The variable ictinvest, being the percentage of ICT in total investments is individually significant, while union and unemp are only jointly significant with ictinvest. All three variables together are jointly significant at the 1% level. Nevertheless, we formally test for multicollinearity, therefore the variance inflation factors (VIFs) are estimated for the independent variables. Table 13 in the appendix displays the results, all tolerance values for the independent variables are slightly above 1, same as the mean VIF, therefore we can conclude that multicollinearity is not present.
Also, the modified Wald test for groupwise heteroscedasticity in fixed effect regression model added to the evidence of heteroscedasticity, therefore robust standard errors should be used. For all three wage bill share regressions we reject the null hypothesis of no first-order autocorrelation of the Wooldrige test (Drukker, 2003) for Autocorrelation in Panel Data (see table 14 in the appendix). Because heteroscedasticity and autocorrelation are present, cluster robust standard errors will be used in all regression estimations.
all tests applied do not let us reject the null hypothesis and therefore random effects seem preferable.
On one hand regressions using fixed effects are advantageous, because countries can be accounted for individually. Therefore, unobservable country characteristics are considered and they allow unobserved effects to be correlated with explanatory variables. The fixed effects regression equation is defined as:
𝐵𝑆𝑆 𝑖𝑡 = 𝛼𝑖+ 𝛽1 𝐼𝐼𝐶𝑇 𝐼𝑇𝑜𝑡 𝑖𝑡 + 𝛽2 𝑈𝑛𝑒𝑚𝑝𝑖𝑡 + 𝛽3 𝑈𝑛𝑖𝑜𝑛𝑖𝑡+ 𝜀𝑖𝑡 [9]
Where the intercept parameters 𝛼𝑖 capture the country specific estimates, the fixed effects, where time-invariant individual characteristics of the country i are included (Hill, Griffiths, and Lim 2012, 543). The coefficients 𝛽1, 𝛽2 and 𝛽3 are constant across countries.
On the other hand, random effects take into account individual, randomized effects and also allow to include time-invariant variables. Then again, individual country specific characteristics are reflected in individual intercept parameters, but these individual differences are not treated as fixed but as random due to the random selection of countries. To reflect these individual randomized effects the intercept parameter consists of a fixed part 𝛼̅ , the population average 𝑖
and a random part accounting for country specific differences 𝑢𝑖, the random effects. Therefore,
the random effects model has two error components combined in 𝑣𝑖𝑡: the random individual error term 𝑢𝑖𝑡 and the regression error term 𝜀𝑖𝑡 (Hill, Griffiths, and Lim 2012, 552).
Our random effects regression equation is defined as follows: 𝐵𝑆𝑆𝑖𝑡 = 𝛼̅ + 𝛽𝑖 1
𝐼𝐼𝐶𝑇 𝐼𝑇𝑜𝑡 𝑖𝑡
+ 𝛽2 𝑈𝑛𝑒𝑚𝑝𝑖𝑡 + 𝛽3 𝑈𝑛𝑖𝑜𝑛𝑖𝑡 + 𝑣𝑖𝑡 [10]
With regards to the differing implications of the specification tests for the regression models, both random and fixed effects will be used in order to provide complete results and to allow comparability of the three regression equations.
4. Econometric Results
further investigate whether the results from the baseline regression are robust, when we only regard Continental Europe and when we omit observations from single countries.
4.1 Basic Results
26
TABLE 6 – FIXED EFFECTS MODEL
High-Skilled Wage Bill Shares Middle-Skilled Wage Bill Shares Low-Skilled Wage Bill Shares
(1) (2) (3) (4) (5) (6) (7) (8) (9)
ICT Intensity 1.172* 1.646** 0.927 0.850 -2.099** -2.497*
(0.386) (0.378) (0.487) (0.696) (0.596) (0.935)
Union Density Rate -0.184 0.500 -0.433 -0.0799 0.617 -0.420
(0.189) (0.240) (0.280) (0.379) (0.426) (0.599) Unemployment Rate 0.255 -0.151 0.0924 -0.117 -0.347 0.268 (0.619) (0.317) (0.527) (0.568) (1.078) (0.831) Constant 15.25 -3.475 -27.42 87.19*** 60.19*** 65.16** -2.441 43.28*** 62.26 (9.281) (4.454) (12.30) (10.88) (5.615) (18.35) (17.68) (6.872) (29.22) Observations 225 225 225 225 225 225 225 225 225 Groups 9 9 9 9 9 9 9 9 9 R2 (within) 0.0332 0.492 0.622 0.0907 0.173 0.179 0.0799 0.405 0.428 F-statistic 0.594 9.208 9.807 1.210 3.621 1.406 1.052 12.40 4.765 p-value (Prob>F) 0.575 0.0162 0.00468 0.347 0.0935 0.310 0.393 0.00783 0.0344 Rho 0.855 0.900 0.963 0.958 0.956 0.955 0.912 0.913 0.921
Coefficients estimated using fixed effects estimation with robust standard errors in parentheses * p < 0.05, ** p < 0.01, *** p < 0.00
TABLE 7 – RANDOM EFFECTS MODEL
High-Skilled Wage Bill Shares Middle-Skilled Wage Bill Shares Low-Skilled Wage Bill Shares
(1) (2) (3) (4) (5) (6) (7) (8) (9)
ICT Intensity 1.170** 1.606*** 0.927 0.864 -2.098*** -2.432**
(0.385) (0.386) (0.487) (0.681) (0.593) (0.874)
Union Density Rate -0.154 0.459* -0.397 -0.0654 0.456 -0.357
(0.191) (0.226) (0.281) (0.362) (0.431) (0.518) Unemployment Rate 0.229 -0.123 0.0356 -0.149 -0.111 0.283 (0.597) (0.304) (0.554) (0.576) (1.133) (0.825) Constant 14.27 -3.454 -25.58* 86.23*** 60.19*** 64.68** 2.013 43.27*** 58.92 (7.936) (2.079) (10.89) (17.93) (11.38) (24.38) (22.25) (11.88) (31.08) Observations 225 225 225 225 225 225 225 225 225 Groups 9 9 9 9 9 9 9 9 9 R2 (within) 0.0331 0.4919 0.6211 0.0903 0.1732 0.1790 0.0773 0.4053 0.4271 Chi2 0.985 9.231 27.94 2.155 3.626 3.999 1.148 12.50 14.07
p-value (Prob > chi2) 0.611 0.00238 0.000004 0.340 0.0569 0.262 0.563 0.000407 0.00281
Rho 0.863 0.910 0.949 0.933 0.961 0.948 0.784 0.923 0.873
Table 6 shows the main results for the fixed effects approach and the three regression equations with the dependent variables high-skilled wage bill share (column 1-3), middle-skilled wage bill share (column 4-6), and the low-skilled wage bill share (7-9). In all three skill group regressions the F-statistic increases and therefore the R2 rises when our main explanatory variable ICT investment / total investment is added to the control variables. While the regression results are insignificant, when only the control variables union density rate and unemployment rate are included, the models solely including the ICT intensity are significant at the 5% level for the high- and low-skilled wage bill shares. Then ICT intensity is positively and significantly correlated with the high-skilled wage bill share. In contrast to that, ICT intensity is negatively correlated with the low-skilled wage bill share. These positive and negative associations become even stronger when we control for unions and unemployment and both the high- and the low-skilled wage bill share fixed effects regressions, including all dependent variables (columns 3 and 9), remain significant at the 5 % level. Moreover, despite adding the two control variables to the wage bill share regressions for each skill group, the increase in the R2 in the columns 6 and 9 is small, indicating that in the middle and low skilled group union and unemployment have a relatively small contribution to the explanatory power of our specification.
Turning to table 7, the random effects results for the three wage bill share equations are provided. We can see that the three regressions including only the control variables union and unemployment are insignificant at the 5% level, like in the fixed effects approach. Again the wage bill share regression of the middle-skilled remains insignificant, no matter which variables are included. In contrast to the other skill groups, the percentage of explained variation of the dependent variables is considerably low for the middle-skilled wage bill share regressions and also does not increase much when additional explanatory variables are added. The goodness of fit measure R2 for our preferred model specification including all variables for the high- and low-skilled are considerably higher with 62% and 42%, which is also the case in the fixed effects approach. Furthermore, the marginal effects of ICT on the wage bill share regressions of high- and low skilled are only slightly smaller than in the fixed effects approach. The high ρ indicates the interclass correlation and shows that a high fraction of the total variance is attributable to differences between countries, again stressing the heterogeneity between countries. The estimated standard deviation of the individual effect ui is in all regressions more
In the decomposition of the wage bill shares into the relative wages and hours in table 8 below we gain further details on the impact of ICTs on the labor market. The columns 1, 4 and 7 are identical with the above reported results of table 7. Like Michaels, Natraj and van Reenen (2014) and Naticchioni and Ragusa (2014) we therefore aim to investigate the impact of ICT intensity on the relative skill prices (wages) and relative quantities (hours). As previously stated, the results for the fixed and random effects are quite similar. Therefore, only the results from the random effects model for the decomposition of the wage bill share into its components will be reported. Table 17 in the appendix shows the results for the fixed effects model.
As already identified in table 6 and 7, ICT intensity is positively correlated with the high-skilled wage bill share. Now table 8 points out that this is driven by a positive and significant impact of ICT intensity on relative wages and relative hours. The estimated coefficients of the high-skilled group reveal that the impact on the relative wages is much stronger than the impact on relative hours, with coefficients of 1.44 and 1.02. At the same time, the impact of ICT on the dependent variable low-skilled wage bill share seems to be driven almost equally through the impact on relative wages and relative hours, the significant negative marginal effect of ICT on the wage bill shares components is around 1.8. In table 7 we saw that for the middle-wage group the regressors were jointly insignificant displayed in the low values for the F-statistics. In table 8 the results for the middle-skilled groups (columns 4 and 5) do not present joint-significance. However, the regression estimation in column 6 with the dependent variable relative hours
High-Skilled Middle-Skilled Low-Skilled
(1) (2) (3) (4) (5) (6) (7) (8) (9) BSH 𝑊 𝐻 𝑊𝑇𝑜𝑡 ⁄ 𝐻𝐻 𝐻𝑇𝑜𝑡𝑎𝑙 ⁄ BSM 𝑊𝑀 𝑊𝑇𝑜𝑡 ⁄ 𝐻𝑀 𝐻𝑇𝑜𝑡 ⁄ BSL 𝑊𝐿 𝑊𝑇𝑜𝑡 ⁄ 𝐻𝐿⁄𝐻𝑇𝑜𝑡 ICT Intensity 1.606*** 1.438*** 1.018*** 0.864 0.382 0.849* -2.43** -1.807*** -1.864*** (0.386) (0.252) (0.249) (0.681) (0.364) (0.387) (0.874) (0.496) (0.566) Union Density Rate 0.459* 0.0317 0.0525 -0.0654 0.0240 -0.0834 -0.357 -0.0462 0.0314 (0.226) (0.167) (0.164) (0.362) (0.215) (0.259) (0.518) (0.338) (0.389) Unemploy-ment Rate -0.123 -0.291 -0.0399 -0.149 0.0538 -0.152 0.283 0.246 0.203 (0.304) (0.192) (0.195) (0.576) (0.355) (0.351) (0.825) (0.477) (0.531) Constant -25.58* 8.057 0.586 64.68** 50.12*** 54.19*** 58.92 41.23* 45.08* (10.89) (6.392) (7.152) (24.38) (13.92) (16.33) (31.08) (18.73) (21.53) Observations 225 225 225 225 225 225 225 225 225 Groups 9 9 9 9 9 9 9 9 9 R2 (within) 0.6211 0.7387 0.657 0.1790 0.088 0.363 0.4271 0.550 0.541 Wald chi2 27.94 133.62 76.70 3.999 5.23 14.88 14.07 39.89 32.35 p-value (Prob > chi2) 0.000004 0.0000 0.0000 0.262 0.1559 0.0019 0.00281 0.0000 0.0000 Rho 0.949 0.965 0.960 0.948 0.974 0.972 0.873 0.904 0.925
Coefficients estimated using random effects estimation with robust standard errors in parentheses
* p < 0.05, ** p < 0.01, *** p < 0.001
worked is significant and the marginal effect of ICT intensity on the relative hours worked is significant at the 5% significance level. Thus, when the ICT intensity increases by 1 percentage point the share of hours performed by the middle-skilled will increase by 0.864 percentage points, ceteris paribus. The effect on the relative wages is comparatively low and remains insignificant, same as for the middle-skilled wage bill share. To put it into contrast, the results of the decomposition of Michaels, Natraj and van Reenen (2014) show that the adjustment in wages and hours was more or less similar, while Naticchioni and Ragusa (2014) found that the impact on the skill groups wage bill shares was driven by changes in hours worked. While results are more mixed, the impact on the high-skilled seems to be more driven by the impact of ICT intensity on wages, the impact on the middle-skilled more by the impact on hours, while the impact on the components of the low-skilled wage bill share is more balanced.
To summarize, the basic results of both the fixed and the random effects approach indicate a positive and significant impact of ICT intensity of the high-skilled wage bill share, which is in accordance with the routinization hypothesis. Thus, ICTs seems to be complementary to the analytical tasks performed by the high-skilled and we cannot reject our first hypothesis. Even though the results for the middle-skilled wage bill share regressions remain insignificant for all specifications, table 8 indicated that ICT positively impacts the middle-skilled through a positive employment effect (column 6). After all the marginal effect on the hours performed by the skilled, the effect is comparatively lower than the employment effect in the high-skilled group. Additionally, the ICT intensity coefficient of the middle-high-skilled wage bill share regression is lower than in the high-skilled wage bill share regression, even though it is insignificant. However, we reject our second hypothesis. Yet, with regards to the low-skilled group and their wage bill share, we cannot find support for the routinization hypothesis predicting no impact of ICT intensity on the low-skilled wage bill share. Because the coefficients of ICT intensity in the low-skilled wage bill share regressions are negative and significant at the 1%-level, we can reject our third hypothesis.
4.2 Sensitivity Analysis
𝐵𝑆𝑆𝑖𝑡 = 𝛼 + 𝛽1 𝐼
𝐼𝐶𝑇
𝐼𝑇𝑜𝑡 𝑖𝑡
+ 𝛽2 𝑈𝑆𝑖𝑡 + +𝛽3 𝑈𝑛𝑖𝑜𝑛𝑖𝑡 + 𝜀𝑖𝑡 [8]
Due to the availability of data we use a smaller sample of 7 countries, without USA and Japan for the time-period of 1996 to 2004. The number of observations of the 7 country groups therefore declines to 62. We include the unemployment by skill group according to the educational level. The time-series plots for the unemployment rates by skill level in figure 3 reveal that unemployment rates are the highest in the low-skilled group across country and years.
The highest levels of the low-skilled unemployment rate can be found in Spain and Finland with over 22%, even though the levels decreased in the following 8 years, the levels remain the highest in the cross-country comparison. Looking at the middle-skilled unemployment rate we can see that also in this educational group the levels are the highest in Spain and Finland over the whole period. The unemployment rate of the high-skilled group is relatively low in most countries, with levels below 10% in most countries and a mean of under 5%, again Spain presents an exception from the rule.
Table 9 reports the results for the regression specification [9] including the unemployment by skill type with a fixed and random effects approach, they can be compared to columns 3, 6, and 9 in table 6 and 7, where ICT intensity has a positive and significant impact on the high-skilled wage bill share.
Now, when we control for the unemployment by skill level, the results remain significant but the marginal effect of ICT intensity changes its sign. This result is at odds with the skill-biased technological change hypothesis and our hypothesis number 1. The constant in our base model has a negative value of 27 in the fixed effects model, now the constant increased to a positive value of 34. The impact of the control union density rate and the newly introduced control unemployment by skill type are highly significant, with a negative association of union density rate and an even stronger negative association of unemployment of the high-skilled group with their wage bill share. Union Density rate in this panel has a positive association with middle- and low-skilled wage bill shares, even though not significant at conventional levels, and a negative association with the bill share of the high-skilled, illustrating an inequality reducing effect of union density rate over the regarded period and when we control for unemployment by skill level. The results for the key variable ICT intensity for the middle-skilled wage bill share regressions have the same sign as in the baseline regressions only the marginal effects of ICTs have decreased in their magnitude. Also, the constant has not changed in the new specification of the middle-bill share regression.
At first view these results imply that the robustness of the previous estimations is not given when we use the new measure of unemployment, the unemployment by skill level. However, when we run our baseline regression for the same time-period of 1996 to 2004, the estimates
TABLE 9 – FIXED AND RANDOM EFFECTS MODEL WITH UNEMPLOYMENT BY SKILL LEVEL
High-Skilled Wage Bill Share
Middle-Skilled Wage Bill Share
Low-Skilled Wage Bill Share
FE RE FE RE FE RE
ICT Intensity -0.248* -0.231* 0.189 0.193 0.263 0.191
(0.0898) (0.0946) (0.235) (0.228) (0.395) (0.343)
Union Density Rate -0.402** -0.375*** 0.175 0.168 0.392 0.303
(0.0700) (0.0743) (0.188) (0.183) (0.220) (0.205) High-Skilled Unemployment Rate -0.606*** -0.617*** (0.0576) (0.0526) Middle-Skilled Unemployment Rate -0.550** -0.555*** (0.134) (0.152) Low-Skilled Unemployment Rate 0.790 0.828* (0.376) (0.389) Constant 34.73*** 33.34*** 68.64*** 69.15*** -14.45 -10.32 (3.901) (7.330) (8.453) (9.355) (12.40) (11.97) Observations 62 62 62 62 62 62 Groups 7 7 7 7 7 7 R2 (within) 0.787 0.786 0.417 0.417 0.582 0.579 F-statistic 99.36 9.833 3.136 p-value (Prob>F) 0.0000 0.0099 0.109 Wald chi2 502.50 18.24 11.08 Prob > chi2 0.0000 0.0004 0.0113 Rho 0.999 0.998 0.997 0.996 0.995 0.993
Coefficients estimated using fixed effects and random effects estimations with robust standard errors in parentheses
