To what extent have changes in the ICT capital stock from
2000-2014 altered the aggregate demand for labour across
industries? - An analysis of 7 Eurozone countries.
Name:
Willem van Lindonk
Student number:
10615091
Supervisor:
Melvin Vooren
ABSTRACT
Concerns about increasing advances in computer science over the past decades have led economists and policymakers to dust off their classical economics literature and recite Keynes’ notion of technological unemployment (1930, p.360). In recent literature it has been shown that computers affect the workplace through altering the demand for skills and changing the returns to different types of labour, implying that computers exhibit both a substitutional as well as a complementary effect on labour demand. In this paper an attempt is made at estimating the aggregate change in labour demand attributable to changes in ICT-capital intensity over a period of 15 years in 7 European countries by using data from the EuKLEMS database, in order to discover whether a substitution or net-complement effect has been at play.
Keywords: Automation, Capital, Complementarity, Computerization, EUKLEMS, ICT, Labour, Technology.
June 2017
University of Amsterdam
Statement of Originality
This document is written by Willem van Lindonk who declares to take full responsibility for the contents of this document.
I declare that the text and the work presented in this document is original and that no sources other than those mentioned in the text and its references have been used in creating it.
The Faculty of Economics and Business is responsible solely for the supervision of completion of the work, not for the contents.
Table of contents
1. Introduction
2. A very brief history of workplace automation
3. Survey of related literature
4. Theoretical framework and method
5. Data
6. Results
6.1 Panel Diagnostics
6.2 Part I: 15-Year Sample (2000-2014) excluding education
6.2.1 Diagnosis6.2.2 Regression Results
6.3 Part II: 7-year sample (2008-2014), including education
6.3.1 Diagnosis6.3.2 Regression Results
7. Discussion
8. Conclusion
1. Introduction
Due to decreasing costs of computer-processing-power (Nordhaus, 2007, p.144) and
continuous development of machine learning and artificial intelligence, more and more jobs are at risk of being taken over by computers (Frey and Osborne, 2017, p. 266). This means that in the coming decades it is very likely that there may be a significant share of the general population losing its job to automation. It is still the question to what extent these jobs will be replaced by new jobs through job-creation and job-diversion to other professions in which demand may increase as a result of increased automation (Autor, 2015, p.3).
Economic models of production such as the Cobb-Douglas model and the more general Constant Elasticity of Substitution model, assume a certain degree of substitutability between capital and labour. Some technologies may give capital a substituting effect for labour whilst other
technologies may lead to a complementary effect to labour. The widespread notion of “Technological Unemployment” noted by Keynes in the 1930’s in his essay “Economic possibilities for our
grandchildren” (1930, p.360) takes the view that this substitution of capital for labour will in the future turn into a situation of close to perfect substitutes of capital for labour, in which the relative cost of each of the factors of production will determine which one will be used and which one not.
The increased use of smarter-becoming computers seems to merit such an idea. If computers take over humans in more and more domains of labour, which they do (Autor, 2015, p.22), then it seems logical that labour and capital may become close to perfect substitutes in the future. If the cost of this labour-substituting computer capital will be lower than the cost of labour, increased automation will drastically reduce the amount of jobs in the economy and hence the returns to labour, which is not an unlikely scenario because of decreasing costs of computer capital (Nordhaus, 2007, p.144) and increasing costs of labour - due to unions, minimum wage rules et cetera.
In such a, yet, hypothetical situation, most of the returns of production may fall on the holders of capital inputs and not on the holders of labour inputs. A computerized society might therefore become very unequal since a lot of people do not have significant amounts of capital inputs at their disposal which makes them dependent on labour income only. Of course, one of the many benefits of a computerized society is that we as humans are able, for the first time in the known history of all species, to free ourselves from the never-ending work that we need to perform to survive, because we have intelligent machines that can do this work for us. As noted by Autor (2015, p. 28), a move into a situation in which computer capital renders human labour superfluous, changes our main economic problem from a problem of scarcity to a problem of distribution, because it is not a loss of welfare that we will have to be afraid of, but an unequal distribution of it.
It is of course hard to predict whether such a situation will occur and whether human labour will be rendered useless in the process of production. Throughout history humans have been very creative in coming up with new tasks that substitute for older ones that have been taken over by machines, as will be shown in the next paragraph about the history of automation. However, based on past data it is possible to measure whether current advances in technology and increased uptake of ICT-capital have altered demand for employment.
The purpose of this paper is to study the extent to which changes in the value of computer capital employed in production have altered the demand for labour in 7 countries in the eurozone over the years 2000-2014, in order to assess whether increased uptake of computers has so far had a significant effect on labour demand and also, provided that a significant effect is found, whether it has been complementary to or substituting for labour.
The rest of the paper will be structured as follows. First a brief overview of the history of workplace automation will be provided to illustrate how new technologies have shaped labour demand in the past. Secondly an overview of the relevant literature will be provided and an
explanation will be given of how this paper relates to that literature. The paragraph that follows will provide the theoretical framework and method with which this research will be conducted. In the fifth paragraph a few brief remarks are made on how the data is structured, which is followed by a sixth paragraph containing the results of this research. Lastly the results will be discussed in paragraph seven, after which a conclusion will be reached in paragraph eight.
2. A very brief history of workplace automation
It is hard to define when the process of workplace automation has started. It was at the dawn of the industrial revolution - at the end of the 18th to the beginning of the 19th century - that in increasing amounts machines were used to substitute for human and animal labour in the workplace, a process coined with the term “mechanization” (Bessen, 2012, p.49). It has been shown by Gregory Clark in his book “a farewell to alms” (2007, p.277) that the consequence of this mechanization had the effect of a decrease in the relative wage premium for men, who have on average a comparative advantage in physically demanding tasks when compared to women. This mechanization however has not led to widespread technological unemployment as people were still needed to operate the
machines and do the many tasks for which machines could not substitute (Bessen, 2012, p.49). The result was a labour market substituting away from physical labour to less physically, but more dexterity demanding tasks. Through the substitution of “physical labour” by “machine power” total welfare was improved as a consequence, because the process of mechanization removed the “human” constraint in the physically demanding part of the production process (Clark, 2007, pp.277-278).
Later on in the nineteenth century the introduction of assembly lines had a different effect: By breaking up the production of certain goods and services into a series of easily performable tasks, Goldin and Katz (1998, p.704) argue that the change of artisan production processes to factory
production led to capital and unskilled labour substituting for skilled labour, effectively de-skilling the production process for many goods and some services. It was at the dawn of the 20th century that this process reversed and electricity gradually substituted for processes that were previously steam powered, which enabled the use of more advanced machines that broadened the scope of workplace automation (Goldin and Katz, 1998, p.697). The effect of this change was that machines now substituted for not only the physically demanding tasks, but also for many routine manual tasks, that could previously only be performed by humans. This reduced the demand for unskilled labour and
increased demand for skilled labour, that was needed to operate the machines (Goldin and Katz, 1998, p.713).
Later in the 20th century, the rapid increase of computer use in the labour market after the 1960’s, driven by the use of transistors and microprocessors (Nordhaus, 2007, p.133) happened together with an increasing wage-bill share for high skilled labour between the 1970’s till the 1990’s (Autor, Katz and Krueger, p.1194) and has led to a decrease in demand for middle-educated
manufacturing jobs, with a rise in low skill service jobs and high skill jobs (Autor, Katz and Kearney, 2006, p.191)(Goos and Manning, 2007, pp.118-119)(Goos, Manning and Salomons, 2009, p. 60). This polarization of the labour market has led to several theories regarding the effect of automation on the labour market which will be discussed in the next paragraph. The point of this paragraph is to illustrate that every technological revolution has substituted for human labour in some aspect of the production process, but at the same time has induced people to move into tasks for which they are complements, be it with some delay. The question is, if one of these effects has dominated labour demand over the past years in the countries under study, and if so, which of these effects seems to have dominated.
3. Survey of related literature
It was Zvi Grichilles (1969, p.467) who was one of the first economists to note that capital and skill are relative complements, meaning, that both high-skilled and low-skilled labour are complements to capital, but that high-skilled labour is more complementary than low-skilled labour. This observation and similar ones after Grichilles led to the development of the Skill-Biased-Change Hypothesis, which states that a new burst in technology raises demand for high skilled workers more than for low and medium skilled workers (Card and Dinardo, 2002, p. 734). It has been found by Katz and Murphy (1992, p. 45) and Levy and Murnane (1992, p.1371) that the demand for high-skilled labour with respect to less-skilled labour has increased since the 1980’s, as well as the wage ratio between the lower and higher educated part of the population. Furthermore it has been found by Krueger (1993, p.49) that a lot of the resulting wage variation can be explained by increased computer uptake amongst highly educated individuals.
At the end of the 1990’s and in the early 2000’s however, as noted by Card and Dinardo (2002, p.771), there were still some puzzles confronting the Skill-Biased-Technological Change hypothesis: increases in demand for high skilled labour were not as large as the hypothesis predicted and decreases in demand for low skilled labour were also not prominent. In a seminal paper by Autor, Levy and Murnane (2003, p.1280) it was noted that the Skill-Biased-Technical change hypothesis fails to explain what it is that computers do, that increases the returns to high skilled labour with respect to lower skilled labour. In their paper they find that computers decrease the returns to routine cognitive and manual tasks, by substituting for them, and increase the returns to non-routine analytic and interactive tasks, by complementing these tasks (p.1308). The task-based model used by Autor, Levy and Murnane provides a more nuanced picture than the Skill-Biased-Technological Change hypothesis of how computers have moved the demand for labour away from routine tasks towards
non-routine tasks (p.1318). It is through these findings that the use of the term Routine-Biased-Technological Change seems more suitable to describe the extent to which computer capital currently substitutes for human labour.When it comes to the distribution of routine and non-routine tasks among the working population, it is found that especially jobs in the middle of the skill distribution contain many routine tasks that can be easily taken over by computers, while in the lower and higher part of the skill distribution the amount of routine task intensity is lower (Autor, Katz and Kearney, 2006, p. 193)(Michaels, Natraj and Van Reenen, 2014, p.60). The result of these differences in routine tasks between occupations has led to polarization of the labour market, for which evidence has been found both in the United States (Autor, Katz and Kearney, 2006, p.191)(Autor, Katz and Kearney, 2008, p. 309)(Autor and Dorn, 2013, p.1574), in the United Kingdom (Goos and Manning, 2007, p.127), in Germany (Spitz-Oener, 2006, pp.244-245)(Dustmann, Ludsteck and Schonberg, 2009, pp. 857-858) and in comparisons of European countries (Goos, Manning and Salomons, 2009, p.61)(Goos,
Manning and Salomons, 2014, p.2515) and a small sample of OECD countries (Michaels, Natraj and Van Reenen, 2014, pp.70-72).
The two papers most similar to this paper are those of Michaels, Natraj and Van Reenen (2014) and Goos, Manning and Salomons (2014). Some of the methodology in this paper is also inspired by these papers. Both papers present clear and strong evidence of employment polarization in European countries using the EUKLEMS database (Michaels, Natraj and Van Reenen, 2014, p.62) and the OECD STAN-Database (Goos, Manning and Salomons, 2014, p.2513). In both of these papers this has been done through selecting a group of High-, Middle- and Low-skilled occupations and regressing changes in these variables as a consequence of respectively ICT capital services (Michaels et al., 2014, p. 61) and a Routine Task Intensity index (Goos et al., 2014, p.2511) to measure respectively the effect on wage dispersion (Michaels et al., 2014, p. 70) and employment share changes (Goos et al., 2014, p.2522).
In this paper the EUKLEMS database (also see Jäger, 2016, pp.1-2) will be used to assess 1 the effect of ICT-capital on labour in a more aggregated manner than in the paper by Michaels, Natraj and Van Reenen. The EUKLEMS database is a project that has been funded by the European
Commission to “create a database on measures of economic growth, productivity, employment creation, capital formation and technological change at the industry level for all European Union member states from 1970 onwards.” Data from the database have been obtained and derived from 2 national statistics offices and Eurostat using the growth accounting method explained by O'Mahony, Timmer, and Van Ark (2007, pp.65-67). Furthermore the data have been classified according to ESA 2010 and ISIC Rev.4 classifications (Jäger, 2016, pp.2-3). 3 4
Obtained from EuKLEMS: http://www.euklems.net/index.html (accessed on 19-06-2017)
1
Obtained from EuKLEMS: http://www.euklems.net/project_site.html (accessed on 19-06-2017)
2
Obtained from European Commision: http://ec.europa.eu/eurostat/web/esa-2010 (accessed on 19-06-2017)
3
Obtained from UNstats: https://unstats.un.org/unsd/cr/registry/isic-4.asp (accessed on 19-06-2017)
The industry level aggregates of 7 eurozone countries from 2000-2014 will be used. The countries that will be investigated are Austria, Finland, France, Germany, Italy, the Netherlands and Spain. Together these countries account for approximately 84% of the population (285.2 million out of 340 million people) and for approximately 88% (9409 out of 10745 billion euros) of all GDP in 5 6 the Eurozone. Two other non-Eurozone countries for which data is available are Sweden and The United Kingdom. However, these countries have been excluded from the analysis due to the fact that in both countries another currency is used than the euro and data have been expressed in national currencies. Exchange rate conversion through matching exchange rates at a certain date and time to yearly observations is tricky and may bias the estimates, because exchange rates may change within a given year and may not be a good representation of the true value of capital applied in a certain year as well as of the value created within a particular industry.
The difference in approach in this paper lies in the manner in which the effect of ICT capital is measured. Michaels et al. (2014, p.70) and Goos et al. (2014, p.2522) take an approach in which they select a number of occupation groups in high, medium and low education categories. The results of these analyses corroborated the Routine-Biased-Technological-Change hypothesis and the
polarization that occurred as a consequence of it. The choice to include certain occupation groups is necessary to be able to measure polarization. However, the choice of occupations for such a study is despite its underlying motivations always a bit arbitrary. Furthermore, it is not the purpose of this paper to prove technology induced polarization, since this has been proven to a reasonable extent in the literature mentioned earlier in this paragraph. The purpose of this paper is to measure the effect of ICT-capital stock changes on total labour demand at the industry level, by integrating other types of non-ICT capital and related variables to isolate the effect and see whether it has been complementing or substituting for labour on average. Instead of measuring changes for occupation groups, it will be changes in industry aggregates that will be measured, to asses the overall impact.
Furthermore, in the paper by Michaels et al. (2014, p.61) the data that was used measured changes from the 1980’s to 2004. The EUKLEMS dataset that is used in this paper runs from 2000 to 2014, which may give a more recent picture of capital-skill complementarity than the papers
discussed in this paragraph. Lastly, trade is found to play a small role in explaining changes in the structure of employment demand which seems to disappear when conditioning on technology variables (Michaels et al., 2014, p.73), which provides some evidence for the trade-induced technical change hypothesis, put forth by Bloom, Draca and Van Reenen (2016, p.95). It is for this reason a closed-economy model will be considered in this paper, by not including a trade-measuring variable.
Obtained from Eurostat: http://ec.europa.eu/eurostat/tgm/table.do?
5
tab=table&init=1&language=en&pcode=tps00001&plugin=1 (accessed on 19-06-2017)
Obtained from Eurostat: http://appsso.eurostat.ec.europa.eu/nui/show.do?dataset=nama_10_gdp&lang=en
6
4. Theoretical framework and method
The purpose of this paper is to assess whether the increased uptake and use of information and communication technologies has affected the demand for labour through the capital-labour substitution channel. As a proxy for labour demand, the number of hours worked in the economy are used instead of the number of people employed, because people can work on contracts of fewer hours than full-time and this may skew the estimates.
In the model it is assumed that both capital and labour are the only factors of production in the economy and are in some dimensions complements to each other and in other dimensions substitutes to each other. It has been shown by Autor, Levy and Murnane (2003, p.1318) that computer capital substitutes for human labour in certain tasks such as routine manual and cognitive tasks and that computer capital complements human labour in occupations that involve non-routine analytic, cognitive and interaction tasks. In this paper the substitution and complement effects will be aggregated to provide a picture of how the recent growth in the ICT capital stock has affected the aggregate demand for labour regressed at the industry level, in order to assess whether it is the complement effect or the substitution effect of ICT Capital that seems to dominate changes in labour demand.
The paper takes the view of a general macroeconomic cyclical flow model - as shown in figure 1 - in which it is assumed that households have certain amounts of labour and certain amounts of capital at their disposal, that can be employed to generate income, (K> or = 0 and L> or = 0). All the firms in the economy are owned as investment vehicles by these households. Because the
Figure 1
government consumes goods and also invests in the economy using capital and labour and is legally “owned” by its citizens, it will be regarded as just a very large household, that both consumes and produces, and will not be treated separately. Because the focus is on the effect of the ICT capital stock in a certain country, import and export links will not be considered because, as mentioned by
Michaels, Natraj and Van Reenen (2014, p.73), “trade ceases to be important after we condition on technology”. Furthermore, for simplicity purposes it will be assumed that savings by households in a given year equal investments. The capital stocks have determined by the following equation, based on the Solow model (Solow, 1956, p.66):
In which “K” stands for the capital stock in a given industry, “t” denotes the time, “h” denotes the type of capital, “δ” denotes the depreciation rate and “I” denotes investment in the amount of capital. The depreciation rates with which the EUKLEMS database works are assumed to be stable over time and the same for all industries (see table 1). Because different types of capital have different types of uses, it may be assumed that there are different levels of substitutability/complementarity. To distill the effect of ICT capital, different types of capital will be assigned to the groups specified in table 2.
The reasoning behind this categorization is the following: ICT-, R&D- and Equipment Capital affect labour demand in a direct way because these are types of capital that are directly applied with labour in the production process. Examples of such directly used technologies could be: a call center agent using a phone to sell a product, a truck driver using a truck for transporting a load to another location or a student writing his or her bachelor thesis on a computer. Other types of capital are more indirect in the production process and are therefore
expected to provide a less clear effect regarding capital-labour complementarity.
This direct versus non-direct categorization will therefore give a more accurate picture of which part of the variation is explained by which type of capital. All capital variables will be scaled by dividing them by the total number of hours of employment in each industry, to measure the intensity of a certain type of capital in an industry as the value of capital-applied-per-hour. Apart from these four capital variables three other relevant independent variables will be applied in the analysis. The first one will be the amount of Gross Value Added divided by the amount of hours of employment to account for the productivity of labour.
Table 1: Depreciation rates in EUKLEMS database Information
Technologies 31.5%
Communication
Technologies 11.5%
Software and Database
Technologies 31.5% Research and Development Assets 20.0% Transport Equipment 17.0% Other Machinery 12.9% Non-Residential Structures 2.4% Residential Structures 1.1% Other Intellectual Products 12.9% Cultivated Assets 15.1%
The second additional independent variable will be the relative cost of labour over the cost of capital. Because changes in this ratio may alter the demand for labour, this ratio will have to be included to further distill the effect of ICT capital. The third additional variable will be the ratio of high to middle educated employees in a particular industry. This variable is based on the observation that ICT capital has a negative effect on routine tasks and a positive effect on non-routine tasks (Autor, Levy and Murnane, 2003, p.1318). Because of the fact that the middle of the skill distribution contains mostly routine tasks and the top of the distribution mostly non-routine tasks (Michaels, Natraj and Van Reenen, 2014, p.60), it can be expected that for a higher ratio of high to middle educated workers, the complementary effect of ICT capital is stronger, which implies that this ratio should have a positive effect on labour demand. The sum of these variables leads to the following regression functions:
Model 1 (15 year sample excluding education):
Model 2 (7 year sample including education):
Table 2: Capital stock categories Capital Stock Proxies Sub Groups
ICT capital Information Technologies + Communication Technologies + Software and Database Technologies
Research and Development
Capital Research and Development Assets
Equipment Capital Transport Equipment + Other Machinery
Other types of Capital Non-Residential Structures + Residential Structures + Cultivated Assets + Other Intellectual Products
As can be logically inferred, the coefficient of interest in both models is beta 1, the estimator of ICT-intensity. Natural logarithms have been taken of all variables to measure percentage changes over time, because for small changes in the underlying variable, the logarithmic value of the variable is approximately equal to the percentage change in the underlying variable (Stock and Watson, 2015, p.317) and given the fact that most changes within an entity are of small magnitude in percentage terms, it can be assumed that logarithms provide a fairly reasonable approximation. For each sample a
Table 3: Characteristics of model 1 and 2
Data Characteristics Model 1 (15-year Sample) Model 2 (7-year Sample)
Number of industries 29 per country (28 for spain) 17 per country (16 for spain)
Number of countries 7 7
Time frame 15 years (2000-2014) 7 years (2008-2014)
Number of observation
units 202 118
Total number of
observations 3030 826
Table 4: Tabulation of formula components
Component Definition
Empl. Total hours of labour by all persons engaged
Kict Value of ICT-capital stock
Krd Value of RD-capital stock
Kequip Value of Equipment-capital stock
Kocap Value of Other-Capital-Assets-capital stock
Value Added Gross Value Added in production
W Total compensation to labour divided by the number of hours of labour employed
R Total compensation to capital divided by the total value of capital employed
Edu The ratio of the number of highly educated workers over middle educated workers
i Industry of observation
small number of lags of the dependent variable will be included as predictors, because the models turn out to be of a non-stationary nature according to the tests for non-stationarity in tables 7 and 11. Due to data constraints for the education variable, several regressions will be performed with two different samples.
The first sample to be used will be a 15-year dataset (2000-2014) with data on all variables excluding the education variable and contains 3030 observations as noted in table 3. The second sample to be used will be a 7-year dataset (2008-2014) with data on all relevant variables in the model. The sample size of the 7-year dataset will be smaller due to the fact that education statistics have been aggregated at the “Letter-code” level of the ISIC system (see Appendix 2) and therefore the sample contains observations of 17 industry groups per country over a time period of 7 years, which equal 826 observations. Furthermore, because the regression used will be a fixed effects regression with time effects we can assume that all time-invariant entity specific effects (e.g. industry
characteristics, work culture and way of doing business) will be absorbed by the constant and all entity-invariant time specific effects (e.g. business cycle, decreased liquidity due to financial crisis) will be absorbed by the time dummies denoted by the greek “phi” symbol. Lastly, definitions of all other variables in the model can be found in table 4.
5. Data
For some periods for some industries returns to capital and gross value added were negative due to losses in a particular industry, especially around 2008. Because it is impossible to take a logarithm of a negative number, these observations have been labeled as missing. These missing data make the panel unbalanced. Because most missing observations are caused by losses and losses are mostly clustered around 2008, both entity and time-fixed effects will be used in all regressions. It is expected that most of the variation caused by these missing data will be absorbed by the time dummies of the periods in which they occur, because it can be expected that all entities will
experience a similar effect from the financial crisis. Nonetheless it should be taken into account that the estimators may still become biased to some extent due to these missing observations and the missing observations that mostly randomly seem to occur in some of the other variables.
Furthermore, data for industries T and U (see Appendix 2) have been deleted from the list of observations due to scarce data availability. On average these industry groups represent approximately 0.3% of total employment, therefore it is expected that this will not significantly affect the estimates. Lastly, for Spain the set of industry units in the 15-year sample is 28 instead of 29 because industry groups 90-93 and 94-96 are only available as an aggregate and in the 7-year sample the number of industry groups for spain is 16 instead of 17 because data of industry groups R and S was only available in the aggregate.
6. Results
6.1 Panel Diagnostics
In order for the regression to be unbiased, consistent and efficient the following conditions will need to be satisfied or diagnosed and treated:
- Large Outliers are unlikely - No Perfect multicollinearity - Zero conditional mean
- Statistical independence of the errors (no cross-sectional dependence) - Homoskedasticity
- Normality of the error distribution - The data is stationary
6.2 Part I: 15-Year Sample (2000-2014) excluding education
6.2.1 Diagnosis
It can be seen from the summary statistics in table 5 that large outliers are unlikely, given the fact that the kurtosis of each variable is not equal to 0 and smaller than infinity. Furthermore from the correlation table in table 6 it can be deduced that there is no sign of either perfect or imperfect multicollinearity. When testing the full model with all relevant independent variables it is found that there is heteroskedasticity present in the data as shown in line 1 of table 7 by the significant value of the Breusch-Pagan test, so Heteroskedasticity and Autocorrelation-Consistent (HAC) standard errors will have to be used to estimate a robust model.
Furthermore, a Pesaran test has been performed to test for cross-sectional dependence. Due to the fact that there are 76 out of 3030 observations missing for the LN(wage to capital cost ratio) variable it was not possible to calculate the Pesaran statistic with the full model. Therefore the model has been tested without this variable, of which the result has been shown in lines 2a and 2b of table 7, the consequence of this is that if the null hypothesis is rejected in the data it is very likely that this will not be changed by the inclusion of LN(wage to capital cost ratio), however when the null-hypothesis is not rejected this is more likely to be due to the exclusion of this variable. Because a restricted sample is used, non-rejection of the null-hypothesis becomes less credible.
In the test of the sample there appears to be a strong case for cross-sectional dependence in the data, since the p-value of the Pesaran test is approximately 0 (see line 2a of table 7), which can be expected since changes in revenues and employment in a certain industry may have effects on revenues and employment in other industries according to a general equilibrium model and all firms experience the effects of the business cycle. The test becomes insignificant when time effects are included in the regression as shown in line 2b in table 7, which demonstrates the relevance of the inclusion of time dummies in the regression. Line 3 of table 7 shows the value of the Harris-Tzavalis
Table 5: Summary Statistics for 15-year sample
Variable N Mean Std.
Deviation Skewness Kurtosis Smallest Largest
LN (Hours Worked) 3030 5.89 1.56 -0.38 2.87 0.63 9.06 LN (ICT-Intensity) 3030 0.98 1.25 0.20 4.24 -3.48 5.79 LN (R&D-Intensity) 3030 -1.13 9.09 -4.22 19.92 -45.46 5.13 LN (Equipment-Intensity) 3030 2.83 1.22 0.47 3.08 -0.57 7.08 LN (Other Capital-Intensity) 3022 3.62 1.56 1.13 6.08 -5.76 9.12 LN (Value Added per Hour) 3028 3.76 0.72 1.33 5.81 1.69 7.38 LN (Wage to Capital Cost Ratio) 2952 5.04 0.84 0.61 4.51 2.54 10.84
Table 6: Correlation Table for 15-year sample
Variable LN
(ICT-Intensity) LN (R&D-Intensity) (Equipment-LN Intensity) LN (Other Capital-Intensity) LN (Value Added per Hour) LN (Wage to Capital Cost Ratio) LN (ICT-Intensity) 1.0000 X X X X X LN (R&D-Intensity) 0.1926 1.0000 X X X X LN (Equipment-Intensity) 0.3079 0.1851 1.0000 X X X LN (Other Capital-Intensity) 0.1730 -0.1564 0.3358 1.0000 X X
LN (Value Added per
Hour) 0.5473 -0.0670 0.4608 0.6488 1.0000 X
LN (Wage to Capital
Table 7: Summary of Diagnostic Tests for 15-year sample
Diagnostic Test Hypothesis Test statistic (P-value)
1) Breusch-Pagan test
for heteroskedasticity Ho: Constant variance of the error term
𝞆
2 = 35.08 (p = 0.0000) 2a) Pesaran test forcross-sectional
dependence (excluding time dummies)
Ho: There is no cross-sectional
dependence CD = 27.16 (p = 0.0000)
2b) Pesaran test for cross-sectional
dependence (including time dummies)
Ho: There is no cross-sectional
dependence CD = 1.26 (p = 0.2078)
3) Harris-Tzavalis unit-root test for non-stationarity
Ho: Panels contain unit roots
⍴
= 0.7299 (p = 1.0000)4) Hausman test for
endogeneity Ho: Difference in coefficients is not systematic
𝞆
2 = 2726.85 (p = 0.0000)5) Jacques-Bera Test for normality of the residual
Ho: Normality in error distribution LM = 29.33 (p = 0.0000)
6) Ramsey-Reset test for zero conditional mean
Ho: Model has no omitted variables F = 2.14 (p = 0.0929)
7a) F-test for comparing full 1 period lag model with no lag model.
Ho: L1.lnemp = 0 F = 613,890.93 (p = 0.0000)
7b) F-test for comparing full 2 period lag model with 1 lag model.
Ho: L2.lnemp = 0 F = 632.50 (p = 0.0000)
7c) F-test for comparing full 3 period lag model with 2 lag model.
unit-root test for non-stationarity in which the null-hypothesis is not rejected, meaning that the data is of a non-stationary nature. For this reason lags of the dependent variable will be included. In line 4 of table 7 the result of the Hausman test has been shown, which seems to imply that the fixed effects model is the most appropriate model to use in this case.
Furthermore in line 5 of table 7 the result of a Jacques-Berra test for normality of the residual has been shown, which seem to be significant implying non-normality of the residuals and in line 6 the result of the Ramsey Reset test is shown with a p-value just significant at 10%, therefore only weakly discarding zero conditional mean dependence. These two results together imply that there may be reason to believe that there is omitted variable bias, which may cause the estimates to be biased due to correlation with the error term. It is possible that this bias may disappear with the inclusion of an education variable in the 7-year sample, assuming that education is an important determining variable of education and capital-skill complementarity. However, the 7-year sample is likely to have less explanatory power, since the number of observations is approximately one fourth of the 15-year sample. Lastly a test has been performed to determine how many lags to include in the regression, by means of including one at a time and testing the unrestricted model R-squared value to the R-squared value of the restricted model without the lag, as can be seen in lines 7.1, 7.2 and 7.3. It is found that 2 period lags of the dependent variable should be included in the regression.
6.2.2 Regression Results
In the first regression (see table 8) only the lagged values of labour demand are taken as explanatory variables, together with the time dummies. The 1 period lagged variable is significant at the 1% level with a positive effect of 1.13% on employment and the 2 period lagged variable is also significant at the 1% level with a negative effect of -0.21% on employment. In the remaining regressions both lagged variables remain significant at the 1% level and only decrease a very small amount in magnitude. In the second regression the ICT capital level is included and is estimated to be negative with an effect of -0.028%, thus meaning that an increase of one euro in the average applied ICT-capital intensity per hour, results in a decrease of -0.028% of hours worked. In the third
regression other variables that could be thought of as “working together” with ICT capital are included. For instance, in the field of research and development ICT capital is a principal component of the development process. Furthermore many types of equipment are enhanced with ICT
technologies to make them more productive. The effects of this third regression are significant at the 1% level for ICT and equipment capital and insignificant at the 10% level for R&D capital. The estimated effects for equipment and R&D capital respectively are -0.033% and 0.0005% and the effect of ICT capital on employment demand decreases in magnitude from -0.028% to -0.018%. In the fourth regression all types of capital are included. The effects of R&D and ICT capital stay roughly the same and the effect of equipment capital decreases in magnitude to a small extent, while the significance levels are similar to those in regression 3. The other-capital-assets variable is estimated to have a negative effect of -0.013% on employment demand, significant at the 5% level.
The fifth regression takes into account the effect of ICT capital and other capital assets, excluding R&D and equipment capital. Which leads the ICT-capital coefficient to increase in
magnitude to 0.027% (significant at 1%) which may be due to the fact that the effects of R&D capital and equipment capital have been neglected. The other-capital-assets variable increases only to a minor extent in magnitude to 0.015% and stays significant at 5%. The sixth regression entails the inclusion of only ICT and non-capital variables - Gross value added per hour worked and the wage to rent ratio. Due to missing data in the variable ln(wage-to-capital cost ratio) caused by negative returns to capital in some periods for some industries, the number of observations is lower than in the other regressions, which might create a degree of bias in the estimators. The estimates for the value-added variable and the wage-ratio variable are respectively -0.044% and -0.026% (both significant at 1%) for which a possible explanation might be that increases in labour productivity - value added per hour
Table 8: Regression results of 15-year sample
Dependent variable: LN(Hours Worked) Variables (1) (2) (3) (4) (5) (6) (7) LN (ICT-Intensity) -0.0275 (0.0066)*** -0.0183 (0.0070)*** -0.0184 (0.0069)*** -0.0265 (0.0068)*** -0.0235 (0.0070)*** -0.0167 (0.0072)** LN (R&D-Intensity) 0.0005 (0.0006) 0.0005 (0.0006) 0.0004 (0.0007) LN (Equipment-Intensity) -0.0330 (0.0112)*** -0.0291 (0.0110)*** -0.0232 (0.0108)** LN (Other Capital-Intensity) -0.0127 (0.0060)** -0.0153 (0.0069)** -0.0099 (0.0055)*
LN (Value Added per Hour) -0.0436
(0.0138)*** -0.0384 (0.0129)***
LN (Wage to Capital Cost Ratio) -0.0264
(0.0055)*** -0.0227 (0.0049)*** L1. LN(Hours Worked) 1.1341 (0.0564)*** 1.1114 (0.0538)*** 1.0932 (0.048)*** 1.0846 (0.0467)*** 1.0982 (0.0514)*** 1.0783 (0.0508)*** 1.0626 (0.0461)*** L2. LN(Hours Worked) -0.2135 (0.0521)*** -0.1966 (0.0498)*** -0.1892 (0.0459)*** -0.1831 (0.0449)*** -0.1881 (0.0477)*** -0.1726 (0.0473)*** -0.1671 (0.0442)*** Constant 0.4615 (0.0576)*** 0.5152 (0.0575)*** 0.6638 (0.0671)*** 0.7129 (0.0811)*** 0.5958 (0.0728)*** 0.8568 (0.0983)*** 0.9745 (0.1106)*** R-squared Within 0.9052 0.9071 0.9088 0.9100 0.9087 0.9086 0.9106 Sample size 2626 2626 2626 2620 2620 2559 2553
Time dummies have been applied in all regressions. Coefficients are in bold and standard errors in parentheses. 10%, 5% and 1% significance levels are indicated by respectively *, ** and ***.
worked - depresses labour demand, due to the fact that more work can be done by fewer hands, and that the positive effect of wanting to hire more workers because of increased productivity does not change enough to compensate the negative effect. For the wage ratio the negative effect is rather obvious, because increases in the relative cost of labour, decrease relative demand for that factor of production.
In the seventh regression all dependent variables have been included. The effect of ICT capital on employment is depressed to a level similar to those in regressions 3 and 4, with a lower significance (-0,0167%, significant at 5%). Again the missing observations in the wage-to-capital-cost-ratio may cause bias in the estimates, however, the effect of ICT capital in regression 7 does not change a lot with respect to the regressions 3 and 4 which might convey that there is not a lot of bias caused by the missing observations. The effects of the equipment and other-capital-assets variable both decrease in strength to respectively -0.023% and -0.010%, with equipment significant at the 5% level and other capital assets significant at the 10% level. The R&D variable stays approximately the same and insignificant. The value-added variable and the wage-ratio variable also decrease in
magnitude to respectively -0.038% and -0.023%, both significant at the 1% level.
The results of these regressions seem to imply that it can be fairly stated that there is at least a strong relation between changes in the value of ICT capital per hour worked and the demand for labour, be it a small effect. Whether this also implies that ICT capital truly exhibits a substitution effect for aggregate labour demand will be discussed in paragraph 7. In the next paragraph the
education variable will be included in a smaller sample to see whether the results of the regressions on the first sample are similar.
6.3 Part II: 7-year sample (2008-2014), including education
6.3.1 Diagnosis
Due to the fact that the regressions on the 7-year sample will be very much like the regressions on the 15-year sample, the same econometric conditions apply. As can be seen in the summary statistics in table 9 the kurtosis of each variable is in the range of 2.6 to 3.8, meaning that outliers are unlikely. From table 10 it can be seen that there is some degree of multicollinearity present in the model. More specifically, the correlation between ICT capital and value added per hour worked is equal to 0.74, the correlation between other capital assets and value added per hour is 0.73 and there are some relatively high degrees of correlation between R&D capital and ICT capital (0.67), Equipment and ICT capital (0.58) and Equipment capital and value added per hour worked (0.68). This multicollinearity may cause large standard errors in the estimators, which may render them insignificant, and increase the probability of a type 2 error. In line 1 of table 11 a test for
heteroscedasticity has been performed, the result of the test is a p-value of approximately 0 indicating a high degree of heteroskedasticity, therefore the fixed effects model will have to be estimated with robust standard errors.
Table 9: Summary Statistics for 7-year sample
Variable N Mean Std.
Deviation Skewness Kurtosis Smallest Largest
LN (Hours Worked) 826 6.59 1.48 -0.63 3.10 2.29 9.31 LN (ICT-Intensity) 826 1.16 1.56 0.54 3.42 -2.66 5.86 LN (R&D-Intensity) 762 0.30 2.33 0.30 2.92 -4.61 6.12 LN (Equipment-Intensity) 826 2.83 1.54 0.89 3.23 -0.56 7.17 LN (Other Capital-Intensity) 826 4.35 1.71 0.83 3.68 1.11 9.12
LN (Value Added per
Hour) 826 4.03 1.07 1.11 3.30 1.88 7.38
LN (Wage to Capital
Cost Ratio) 801 5.54 1.02 0.26 3.82 2.95 10.86
LN (Education High to
Middle ratio) 826 -0.70 1.09 0.11 2.58 -3.22 2.58
Table 10: Correlation Table for 7-year sample
Variable LN
(ICT-Intensity) LN (R&D-Intensity) (Equipment-LN Intensity) LN (Other Capital-Intensity) LN (Value Added per Hour) LN (Wage to Capital Cost Ratio) LN (Education High to Middle ratio) LN (ICT-Intensity) 1.0000 X X X X X X LN (R&D-Intensity) 0.6661 1.0000 X X X X X LN (Equipment-Intensity) 0.5849 0.5518 1.0000 X X X X LN (Other Capital-Intensity) 0.4397 0.2568 0.5354 1.0000 X X X
LN (Value Added per
Hour) 0.7400 0.5637 0.6770 0.7267 1.0000 X X
LN (Wage to Capital
Cost Ratio) 0.3728 0.4340 0.4248 0.5291 0.4393 1.0000 X
LN (Education High
Table 11: Summary of Diagnostic Tests for 7-year sample
Diagnostic Test Hypothesis Test statistic (P-value)
1) Breusch-Pagan test
for heteroskedasticity Ho: Constant variance of the error term
𝞆
2 = 47.17 (p = 0.0000) 2a) Pesaran test forcross-sectional
dependence (excluding time dummies)
Ho: There is no cross-sectional
dependence CD = 19.26 (p = 0.0000)
2b) Pesaran test for cross-sectional
dependence (including time dummies)
Ho: There is no cross-sectional
dependence CD = -0.11 (p = 1.000)
3) Harris-Tzavalis unit-root test for non-stationarity
Ho: Panels contain unit roots
⍴
= 0.7659 (p = 1.0000)
4) Hausman test for
endogeneity Ho: Difference in coefficients is not systematic
𝞆
2 = 262.83 (p = 0.0000)5) Jacques-Bera Test for normality of the residual
Ho: Normality in error distribution LM = 33.09 (p = 0.0000)
6) Ramsey-Reset test for zero conditional mean
Ho: Model has no omitted variables F = 1.95 (p = 0.1201)
7a) F-test for comparing full 1 period lag model with no lag model.
Ho: L1.lnemp = 0 F = 1090.92 (p = 0.0000)
7b) F-test for comparing full 2 period lag model with 1 lag model.
In lines 2a and 2b a Pesaran test for cross-sectional dependence has been performed. The circumstances under which this test has been performed are similar to those in sample one; the variable R&D-intensity has been dropped due to too much missing observations (64) for the Pesaran test to be performed. This will have the same implications for the validity of the test as in the
diagnosis paragraph of the 15-year sample. It is found that there is evidence for cross-sectional dependence at the 1% level using a Pesaran test, when excluding time dummies. This cross-sectional dependence disappears when including time dummies, which motivates their use in the regressions to be performed. In line 3 of table 11 a unit root test for non-stationarity has been performed. It is found to be significant at approximately 1.000 indicating that the data is non-stationary. Lags will be
included in order to counter this problem. In line 4 a Hausman test has been performed to test whether the fixed effects model is more suitable than a random effects model. As in the 15-year sample the test is significant, implying that for the 7-year sample the fixed effects model is most appropriate.
A Jacques-Berra test for normality of the residuals has been performed in line 5 to test whether the error term is normally distributed. The test comes out significant, indicating that it has to be taken into account that the estimators correlate with the error term, meaning that the estimates that have been found may be biased to some extent. When it comes to omitted variables bias a test for a zero conditional mean of the residuals is performed through performing a Ramsey-Reset test in line 7. The result of the test comes out insignificant and therefore the zero conditional mean assumption seems to be satisfied. Lastly in lines 7a and 7b a model with a one-period lag and a two-period lag have respectively been compared with a model without a lag and a one-period lag of the dependent variable. In summary the results indicate that 1 lag of the dependent variable will have to be included in the regression, time dummies will have to be used to counter cross-sectional dependence and a fixed effects model with heteroskedasticity and autocorrelation robust standard errors will have to be estimated. Furthermore, it will have to be taken account of that multicollinearity may render the estimators insignificant by increasing the standard errors and that the non-normality of the error term may cause some bias in the estimates.
6.3.2 Regression Results
First a regression on employment is performed (as can be seen in table 12) by using a 1 period lag of employment as a predictor together with time dummies. The result of the lag is significant at the 1% level with a value of +0.789% per year, which slightly decreases in magnitude through all the regressions, but stays significant at the 1% level. In the second regression ICT capital is included in the regression which is estimated to have a negative effect of 0.056% per year,
significant at the 1% level. The third regression includes R&D capital and Equipment capital in the equation, which reduces the significance of ICT-capital to being insignificant at the 10% level and a value of -0.034%. The values of R&D capital and Equipment capital are respectively -0.009% and -0.036% and are both also insignificant at the 10% level. One possible reason for this insignificance may be the high degree of multicollinearity between ICT, R&D and equipment capital that increases
Table 12: Regression results of 7-year sample Dependent variable: LN(Hours Worked) Variables (1) (2) (3) (4) (5) (6) (7) (8) (9) (10) LN (ICT-Intensity) -0.0557 (0.0185) *** -0.0334 (0.0223) -0.0482 (0.0198) ** -0.0407 (0.0226) * -0.0474 (0.0184) ** -0.0320 (0.0212) -0.0565 (0.0180) *** -0.0324 (0.0212) -0.0375 (0.0195) * LN (R&D-Intensity) -0.0087 (0.0099) -0.0023 (0.0102) -0.0020 (0.0092) -0.0019 (0.0093) LN (Equipment-Intensity) -0.0364 (0.0235) -0.0090 (0.0228) -0.0013 (0.0220) -0.0015 (0.0220) LN (Other Capital-Intensity) -0.1476 (0.0339) *** -0.1408 (0.0308) *** -0.1561 (0.0311) *** -0.1553 (0.0310) *** -0.1581 (0.0336) ***
LN (Value Added per Hour) -0.0641 (0.0239) *** -0.0863 (0.0222) *** -0.0858 (0.0221) *** -0.0870 (0.0216) ***
LN (Wage to Capital Cost Ratio) -0.0038 (0.0026) -0.0048 (0.0025) * -0.0048 (0.0025) * -0.0039 (0.0023) LN (Education High to Middle ratio) -0.0096 (0.0083) -0.0024 (0.0094) -0.0018 (0.0087) L1. LN(Hours Worked) 0.7890 (0.0307) *** 0.7750 (0.0290) *** 0.7603 (0.0301) *** 0.6843 (0.0316) *** 0.6840 (0.0333) *** 0.7672 (0.0326) *** 0.6648 (0.0341) *** 0.7724 (0.0288) *** 0.6648 (0.0342) *** 0.6654 (0.0335) *** Constant 1.3736 (0.2031) *** 1.5278 (0.1866) *** 1.7027 (0.2181) *** 2.7579 (0.3253) *** 2.7392 (0.3492) *** 1.8506 (0.2355) *** 3.2799 (0l3652) *** 1.5387 (0.1859) *** 3.2727 (0.3632) *** 3.2923 (0.3553) *** R-squared Within 0.7761 0.7849 0.7925 0.8098 0.8127 0.7900 0.8211 0.7858 0.8212 0.8188 Sample size 708 708 655 708 655 686 636 708 636 686
Time dummies have been applied in all regressions. Coefficients are in bold and standard errors in Parentheses. 10%, 5% and 1% significance levels are indicated by respectively *, ** and ***.
the standard errors and renders the estimators insignificant. This is partly supported by the fact that if R&D and equipment capital are removed from the equation and when other estimators and interaction effects are included in further regressions the estimator of ICT capital becomes significant again (see regressions 4, 6, 8 and 10). In regression 4 the other-capital-intensity and the ICT capital variable are used as the only capital asset regressors in the model and ICT capital becomes significant at the 5% level with a value of -0.048%. The introduced OCAP variable is significant at the 1% level with a coefficient of -0.148%. In regression 5 all capital variables have been included as regressors and ICT capital decreases in significance from 5% to 10% with a value of -0.041%. Equipment and R&D capital are again both insignificant at 10% and have coefficients of respectively -0.009% and -0.002%.
In the sixth regression the effect of ICT and the non-capital variables “value added” and the “wage-ratio” have been estimated, in which the ICT variable becomes significant at the 5% level again and is estimated at -0.047%. The value added variable is estimated to be significant at the 1% level with a value of -0.064% and the wage-ratio variable is insignificant at the 10% level with a value of -0.004%. Regression 7 on the 7-year sample consists of the same variables - except for the two-period lag - as regression 7 in table 8 on the 15-year sample. However, in contrast to the results on the 15-year sample, ICT is insignificant at the 10% level with a value of -0.032%. Again a possible cause for this may be the high degree of multicollinearity between ICT and the different capital variables as well as the value-added variable (see table 10). Some evidence for this can be seen in the regressions because the estimates for ICT-capital change a lot in significance by each regression and the variable becomes more significant once a highly correlated variable is dropped (regressions 4, 8 and 10) . Furthermore the sample is almost four times as small as in regression 7 of table eighth which may also contribute to the insignificance of the estimators because the magnitude of the standard errors is negatively related to the sample size (Stock and Watson, 2015, p.121)
In regression 8 of table 12 the number of hours worked is regressed on ICT-capital and the education ratio. The value of ICT-capital turns out to be negative at the 1% level with a value of -0.057% and education is insignificant at the 10% level with a value of -0.010%. It is somewhat strange that the coefficient of the education variable is negative, given the fact that according to the literature it should be positive, since higher education is supposed to be a complement for computer capital. It should be noted that the coefficient is insignificant, implying that there is no aggregate effect on employment demand from higher education, meaning that firms hire whoever is available based on their derived demand. Also, provided that the true effect would be significantly negative, it could be possible that highly educated people create more innovations that reduce the need for labour inputs. The consequence of this could be that industries with a higher ratio of high-to-middle educated people are more innovative and less dependent on labour inputs.
In regression 9 all relevant independent variables have been included and the ICT capital effect becomes insignificant again at the 10% level with a value of -0.032%, as do R&D capital, equipment capital and the education variable. This insignificance may again be partly due to the degree of multicollinearity in the model, which may drive up the standard errors and increase the
probability of a type 2 error, which is further supported by the results of regression 10. However, it may also be the case that variables such as value added and other capital assets are much more important in determining employment demand than ICT-capital, and that the inclusion of these variables just removes the bias in the ICT-variable. Regression 10 estimates the full regression without equipment and R&D capital. ICT becomes significant again at the 10% level with a value of -0.0375%, other-capital assets is significant at the 1% level with a value of -0.158%, value added is significant with a value of -0.087% at the 1% level and the wage to capital cost ratio and education variable are both insignificant at the 10% level with a value of respectively -0.004% and -0.002%.
7. Discussion
Based on the results of table 8 and 12 it can be argued that there is at least a reasonable amount of evidence that increases in ICT-capital intensity exhibit a small negative relation with respect to the amount of hours worked in a particular industry, controlling for all other entity-invariant and time-invariant characteristics and for the variation caused by the selection of other independent variables in the model. This provides some evidence for the existence of a dominating substitution effect of computer capital for labour in the countries under study. The effect on employment from ICT-capital was estimated to be somewhere between -0.028% and -0.017% in the first sample without education and was found to be significant in all regressions at at least 5%. In the second sample with education the effect on employment from ICT-capital was estimated to be somewhere between -0.057% and -0.032% with variations in significance from being significant at 1% to being insignificant at 10%, which may have been caused both by the relatively smaller number of observations and observed degree of multicollinearity as mentioned before.
Despite differences in significance of the estimators, the relationship between ICT-capital and the number of hours worked in an industry seems to be quite small. This may be partly due to the fact that the ICT-capital stock is only a very small amount -somewhere between 1.0% to 3.5% - of the total capital stock in each country as can be seen in figure 2. There is evidence of an increasing trend in the share of ICT-capital across all countries in the dataset except for Italy, which may indicate that at the moment computer technology does not have a big effect on employment, because there is just not enough of it being applied in production processes, but that in the future this effect may become stronger, assuming that the upward trend continues. Also, ICT-capital depreciates at a much faster rate than many other types of capital (see table 1) this may create downward bias in the magnitude of the estimator because much more investment - as can be seen in appendix 3 - is needed to maintain the ICT-capital stock than other types of capital and especially due to declining prices of computer technology over time (Nordhaus, 2007, p.144), declining values of computer capital may overestimate the extent to which computer capital declines in usefulness over time and thus further biases the estimator negatively in terms of magnitude.
Furthermore, it may also well be the case that despite the low share of ICT-capital and fast depreciation rate, there is another factor to consider. Most people work on contracts that are at least
valid for a certain time period or until retirement. It is possible that the computer technologies that may more readily substitute for workers are already available to a large extent, however due to labour contracts and other legal constraints that impose obligations on the employer to keep its workers on the payroll, there may be a significant lag of employment effects of these new types of capital which may only become visible once more time has passed than the time period that has been analyzed in this study.
Regarding the internal validity of this study there is one factor that is important to mention here. As can be seen in tables 7 and 10 in the Ramsey-Reset test, the zero conditional mean
assumption was only weakly rejected with a p-value of 9.29% in the first sample, and only non-rejected with a p-value of 12.01% in the second sample. This means that it is not really clear in both samples whether the condition truly applies. If this condition does in essence not apply then the estimators are inconsistent (Stock and Watson, 2015, p.772) which implies that the estimators do not converge to their true value when the sample size tends to infinity and that large numbers of
observations, do not guarantee the validity of the results.
A cause for the possible violation of this condition is most likely to be omitted variables in the regressions. It is very much possible that there are other important non-time-and-entity-invariant variables that have been overlooked in this research that significantly affect both the formation of ICT-capital and the demand for employment. This also entails that the results of this research do no necessarily expose a causal link. Based on our knowledge of what computers can do, which is an expanding range of tasks (Frey and Osborne, 2017, p.255), there are many ways in which computers enter the labour demand equation. This study has been focused on the quantity effects in terms of values and hours worked. However, price composition effects are very important as well, because the
Figure 2: ICT-capital stock as percentage of total capital stock
0,00% 1,00% 2,00% 3,00% 4,00% 2000 2002 2004 2006 2008 2010 2012 2014
Austria Finland France Germany Italy The Netherlands Spain
relative cost of labour to capital is an important variable in determining the factor input choices made by firms. Price effects, however, have only been weakly accounted for by the wage to capital cost ratio, because only data for total labour and capital compensation was available and not specifically for certain types of capital.
Due to the macroeconomic nature of this study it is hard to isolate all factors that lead to computer-labour complementarity/substitution and measure the true aggregate effect on employment. It is therefore up to future research to investigate the extent to which there is truly a dominating substitution effect by using datasets that gather more observations over time and include more relevant dependent variables that have been possibly overlooked in this study.
8. Conclusion
The computer revolution that has started at the end of the 20th century has without a doubt changed the way we work and altered the demand for skills in the workplace. No different than other disruptive technologies in the past has the process of automation, started by the computer revolution, created both losers and winners in different parts of the economy. It has been found in the literature that especially the nature of the tasks performed is an important determinant of how computerization alters demand for certain occupations and hereby changes both the returns and the number of jobs available in different occupational groups.
Evidence has been found for the Routine-Biased-Technological-Change hypothesis, which can explain to some extent why the wage and especially the employment structure seems to be polarizing over the past couple of decades. Most studies that have made an attempt to measure the employment impacts of computerization are of a micro-economic nature. In this paper it has been attempted to use a more macro-economic approach to see whether changes in aggregate labour demand can be attributed to an overall increasing ICT-capital stock and whether this result holds for all industry groups within a number of countries in the European Monetary Union.
Results from the regressions in this paper suggest that there is at least some part of the variation in employment demand attributable to changes in the value of ICT used per hour worked. Estimates are in the range of an effect of approximately -0.056% to -0.016% with a more pronounced, but less significant, negative effect in the smaller sample that includes the education variable and with a less pronounced, but more significant, negative effect in the larger sample that excludes education. These negative estimates imply the possible existence of a net-substitution effect of computer capital for human labour. However, it should be noted that this effect is still of a very small magnitude.
Also, as argued in paragraph 7, some of the diagnostic test results may indicate that there is a possible existence of omitted variables, which is not unlikely given the macro-economic scope of this study. The consequence of this is that it is possible that the estimators are not consistent and do not approach their true values, despite the large number of observations in this study. Further research will have to be done to test whether the results of this study are similar when including other relevant variables that may have been overlooked in this paper in order to provide stronger evidence of a true negative relation between ICT-capital intensity and labour demand.
On a closing note one might ask whether we will all become unemployed at some point in time due to, as Keynes (1930, p.360) so nicely put it, “our discovery of means of economizing the use of labour outrunning the pace at which we can find new uses for labour”. Looking at the current pace of technology one could argue that this is not a very unlikely scenario. As argued by Brynjolfsson and McAfee (2011, pp.13-14) in 2004 it was still widely conceived that driving a car was a skill that no computer would ever be able to perform, due to the complexity of the tasks involved. Six years later, in 2010, Google announced that it had successfully performed tests with automated Toyota Priuses on public roads. Furthermore, due to new programming techniques that create algorithms that infer heuristics and patterns from big data analysis, computing is no longer confined to the formulation of explicit rules. Technologies like these open the door for an increase in machine intelligence, which provides a broader scope of substitution possibilities of ICT capital for human labour. The net effect of substitution will eventually depend on whether we are able to create new jobs that provide added value in this age of increased machine intelligence. For these reasons the topic of human-computer substitution will most likely become even more relevant in the future and as time passes and more data will become available, future research will hopefully get a tighter grasp on the true effect that computer capital exhibits on labour demand.
References:
● Autor, D. (2015). Why Are There Still So Many Jobs? The History and Future of Workplace
Automation. Journal of Economic Perspectives, 29(3), pp.3-30
● Autor, D., Katz, L. and Krueger, A. (1998). Computing inequality- Have computers changed
the labor market? Quarterly Journal of Economics, 113(4), pp.1169-1213.
● Autor, D., Levy, F. and Murnane, R. (2003). The Skill Content of Recent Technological
Change- An Empirical Exploration. Quarterly Journal of Economics, 118 (4), pp.1279-1333.
● Autor, D., Katz, L. and Kearney, M. (2006). The Polarization of the U.S. Labor Market. American Economic Review, 96(2), pp.189-194.
● Autor, D., Levy, F. and Murnane, R. (2007). The Skill Content of Recent Technological
Change- An Empirical Exploration. Quarterly Journal of Economics, 2003, 118. (4), pp.
1279-1333.
● Autor, D., Katz, L. and Kearney, M. (2008). Trends in U.S. Wage Inequality- Revising the
Revisionists. The Review of Economics and Statistics, 90(2), pp.300-323.
● Autor, D. and Dorn, D. (2013). The Growth of Low-Skill Service Jobs and the Polarization of
the US Labor Market. American Economic Review, 103(5), pp.1553-1597
● Bessen, J. (2012). More Machines, Better Machines…or Better Workers, The Journal of Economic History, 72(1), pp.44-74.
● Bloom, N., Draca, M. and Van Reenen, J. (2016). Trade induced technical change - the
impact of Chinese imports on innovation, IT and poductivity. The Review of Economic
Studies, 83(1), pp.87-117.
● Brynjolfsson, E., McAfee, A. (2011). Race against the machine: how the digital revolution is
accelerating innovation, driving productivity, and irreversibly transforming employment and the economy. Digital Frontier Press, Lexington, MA.
● Card, D. and Dinardo, J. (2002). Skill-Biased Technological Change and Rising Wage
Inequality- Some Problems and Puzzles. Journal of Labor Economics, 20(4), pp.733-783.
● Clark, G. (2007). A Farewell to Alms; a brief economic history of the world. New Jersey: Princeton University Press.
● Dustmann, C., Ludsteck, J. and Schonberg, U. (2009). Revisiting the German wage structure. Quarterly Journal of Economics, 124(2), pp.843-881.
● Frey and Osborne. (2017). The future of employment- How susceptible are jobs to
computerisation? Technological Forecasting & Social Change, 114, pp.254-280.
● Goldin, C. and Katz, L. (1998). The Origins of Technology-Skill Complementarity. The Quarterly Journal of Economics, 113(3), pp.693-732.
● Goos, M. and Manning, A. (2007). Lousy and Lovely Jobs- The Rising Polarization of Work
in Britain. The Review of Economics and Statistics, 89(1), pp.118-133.
● Goos, M., Manning, A. and Salomons, A. (2009). Job Polarization in Europe. American Economic Review, 99(2), pp.58-63.
