Automation and Job Polarization: On the Decline of Middling Occupations in Europe

Using data from 10 Western European countries, I provide evidence that the fall of prices of information technologies (IT) is associated with a lower share of employment in middle wage occupations and a higher share of employment in high wage occupations. The decline of IT prices has no robust effect on the share of employment in the lowest paid occupations. Similar results hold within gender, age and education-level groups, with notable differences in these groups. For instance, the share of employment in high wage occupations among females has increased more than among males with the fall of IT prices. This is consistent with arguments that women hold a comparative advantage in communication and social skills, which are in demand in high wage occupations.


Introduction
For quite some time, the consensus has been that most of the recent technological changes have been skill-biased, complementing high-skill workers and substituting for low-skill workers (see, e.g., Katz and Autor, 1999). However, skill-biased technological change alone cannot explain a prominent and relatively recent phenomenon: the decline in the share of middle wage occupations relative to high and low wage occupations. Goos and Manning (2007) call this phenomenon "job polarization." One of the main hypotheses put forward for job polarization is that recent technologies, such as the computers, substitute for routine tasks. These tasks tend to be readily automatable and are usually performed by middle wage occupations, such as stationaryplant operators. They complement nonroutine cognitive tasks, which are usually performed by high wage occupations, such as managers. In turn, the rise of employment in highly paid occupations increases the demand for nonroutine manual tasks, which are usually performed by low wage occupations, such as personal services (see, e.g., Autor, Levy, and Murnane, 2003, Autor and Dorn, 2013, Mazzolari and Ragusa, 2013. In this paper, I empirically investigate the effect of the rapid fall in prices of information technologies (IT) on industries' demand for high, middle and low wage occupations using a difference-in-differences framework in the spirit of Rajan and Zingales (1998).
More specifically, I ask whether the fall in prices of information technologies has affected the demand for high, middle and low wage occupations more in industries which depend more on IT compared to industries which depend less. I use industry-and country-level data from 10 Western European countries and 1993-2007 period to establish the results. I find that the share of employment in middling occupations has declined and the share of employment in high wage occupations has increased with the fall in IT prices. I find no systematic evidence that the fall in IT prices affects the share of employment in the lowest paid occupations. Similar results hold within gender and age groups. These findings provide support for the hypothesis put forward for explaining job polarization.
They are broadly in line with and complement the results of Autor et al. (2003), Autor and Dorn (2013), Acemoglu and Autor (2011), and Goos, Manning, and Salomons (2014), among others. and a growing number of papers offer evidence corroborating this view. Using US data, Autor et al. (2003) show that the use of computers (a type of IT) is associated with reduced employment in middle wage (routine) occupations within industries. Autor and Dorn (2013) show that, in the US, the growth of workplace computer use has been faster in areas which had initially high proportions of routine workers. Goos et al. (2014) show that during the period of 1993-2010 employment has declined in routine task intensive occupations in 16 Western European countries.
The polarization of employment is also mirrored in education-level groups. Acemoglu and Autor (2011) show that in the US the demand for workers with high-and lowlevels of education has increased relative to the demand for workers with medium-level of education. Using data from 11 OECD countries, Michaels, Natraj, and van Reenen (2014) provide evidence that industries with faster growth in information and communication technologies have increased the demand for highly educated workers at the expense of middle-educated, with almost no effect on low-educated workers.
A few recent papers independently explore the differences in the trends of polarization across genders using US data (e.g., Cerina, Moro, andRendall, 2017, Cortes, Jaimovich, andSiu, 2018). Cerina et al. (2017) document that job polarization is more prevalent among females than among males. The results of the current study suggest that the fall in prices of information technologies can be one of the rationales of their finding. 1 The findings of this study complement the results of these papers. An innovation of this study is its identification strategy. I use the assignment of occupations into task/wage groups by Goos et al. (2014) to compute employment in occupation groups with different task contents and utilize a difference-in-differences frameworkà la Rajan and Zingales (1998). In this framework, I employ the variation of IT prices over time and the industrylevel variation of dependence on IT, which allows me to explicitly take into account the technological side of the effect of the fall in IT prices on employment. I provide international evidence corroborating the hypothesis for job polarization. By exploring differences in gender and age groups, I also uncover some more concealed features of the effects of recent technological changes on labor markets in Europe.
The next section describes a simple model to motivate the empirical test. The third section offers the empirical specification, and describes the data and its sources. The fourth section summarizes the results. The last section concludes.

Theoretical Background
A fall in IT prices would increase demand for nonroutine cognitive (abstract) task intensive occupations and reduce demand for routine task intensive occupations more in industries which depend more on information technologies. I present a simple model to show explicitly how such an inference can hold and to set the stage for the empirical analysis. The model bears a resemblance to the models of Acemoglu and Autor (2011) and Autor and Dorn (2013).
The producers use abstract and routine task inputs, T A and T R , and information technologies, IT , to produce homogenous goods, Y . They have a CES production technology, which is given by where α IT > 0, α T R > 0, α ∈ (0, 1), and ε > 1. In the production of Y , α IT measures the relative importance of IT and a higher α IT implies higher share of compensation for IT . In this sense, it measures the technological dependence on IT . In turn, ε is the elasticity of substitution between routine tasks and information technologies, and the elasticity of substitution between abstract tasks and information technologies is equal to 1, by construction. Since ε > 1, information technologies are more complementary to abstract tasks than to routine tasks.
The usual profit maximization implies the following conditions where p IT , p T R , and p T A are the prices of information technologies and task inputs, and the price of Y is normalized to 1.
I assume that p IT , p T R , and p T A are determined at country level. This implies that the derivative of the demand for T A relative to T R with respect to p IT and the change of the absolute value of that derivative with α IT are given by and It is straightforward to show that ∂ ln IT /∂p IT is negative. Therefore, the decline in p IT increases the demand for T A more than the demand for T R when ε > 1, according to equation (5). According to equation (6), in a country, the demand for T A relative to T R would increase more in industries with higher α IT than in industries with lower α IT with the decline in p IT in such a case. This means that T A increases and T R declines with the fall of p IT if employment in Y is fixed and these changes are larger in industries with a larger α IT .
It can be shown that such differential changes can also hold within gender and age groups incorporating these demand functions in a simple Ricardian model of comparative advantage. To do so, I assume that workers are endowed with labor hours, which need to be converted into abstract and routine tasks in order to earn market income. I assume that the conversion function of task k = T A , T R is given by α L,k (u k L) γ ,where α L,k > 0, u k is the share of labor hours L converted to task k, and γ ∈ (0, 1). 2 I normalize α L,T A and set it to equal to 1.
This setup implies that the supply of abstract tasks relative to the supply of routine tasks is given by and the share of employment in abstract tasks is given by It is straightforward to show that in this model economy a fall in p IT increases the share of employment in abstract tasks u T A and it has a stronger effect in industries which have a higher α IT . However, this differential effect of the fall in p IT on u T A (and u T R ) is weaker in groups which have a comparative advantage in converting labor hours into routine tasks (i.e., a higher α L,T R ): The differential changes in T A and T R in industries which depend more on IT than in industries which depend less on IT should be observed in the data as differential changes in the employment in high and medium wage occupations which perform these tasks. I look exactly for such disparities and differential changes in the empirical specification.

Empirical Methodology and Data
where ζ and ξ are country-industry and country-year fixed effects, and η is an error term.
The parameter of interest is β. It is identified from the temporal variation of IT prices, the variation of technological dependence on IT across industries, and within country, time, and industry variation of the interaction term. 3 An advantage of this test is that it alleviates the endogeneity concerns because of omitted country-and industry-level variables. For example, country-industry and country- year fixed effects alleviate the potentially confounding effects of regulatory and discriminatory practices which affect the demand and supply of these tasks. These fixed effects also alleviate the potentially confounding effects of trends in relative wage rates.  Goos et al. (2014), to compute the number of (usual) weekly hours worked in these occupation groups in each sample industry, country, and year. I derive employment shares from the number of hours worked. Goos et al. (2014) also use ELFS data and exclude from the sample some of the occupations and industries because of sample imperfections and potentially large state involvement. These occupations and industries are also excluded from the analysis in this paper. Moreover, similarly to Goos et al. (2014), I use 2-digit aggregation level for occupations throughout the analysis.
Given data availability, the analysis of this paper focuses on 10 Western European countries and the period between 1993 and 2007. The list of sample countries and the sample period for each country are offered in Table 1. and 45), and old (older than 45). I compute the number of hours worked in each of these categories for all occupation group-industry cells in sample countries and years. Table 3 offers basic statistics for the employment shares in high, medium and low wage occupations within each of these categories. The data reveal seemingly intuitive patterns.
On average, men work more in high wage occupations and less in low wage occupations than women, which can contribute to the aggregate wage gap. The share of employment in high wage occupations is higher and the share of employment in low wage occupations is lower among medium-age and old workers than among young workers. The motivation for using data from US industries is that these industries are the world leaders in terms of investments in IT and the level of IT capital. Therefore, the confounding variation in the share of IT capital compensation in industrial value added because of differences in factor input levels is likely to be the smallest in US industries according to equation (2), given that ε > 1. To test this and the validity of this measure, The measure of dependence used in this paper firmly correlates with similar measures used in the literature (see, e.g., Chen, Niebel, and Saam, 2016, Jerbashian and Kochanova, 2016, 2017. I perform a range of robustness checks for it in the Online Appendix -Further Robustness Checks and Results.
to the significant innovations in IT that occurred over the sample years in the US and to the rise of IT production in Asia and, in particular, in China. The country-level variation is likely to be stemming from regulations that affect the access to and adoption of IT.
In turn, the near absence of industry-level variation suggests that the law of one price holds in sample countries. I average the price of investments in IT relative to the price of value added across industries, in sample countries and years, and use that average as the measure of the price of information technologies, p IT .
In the estimations of the baseline specification (10), I use the inverse of this measure. According to the theoretical model, β is then expected to be positive for high wage occupations and negative for medium wage occupations as p IT declines and its inverse increases. This parsimonious theoretical model has no predictions for low wage occupations. Nevertheless, β can be expected to be nil for these occupations since information technologies are not likely to directly affect employment in these occupations (Autor et al., 2003, Autor andDorn, 2013 Table 2 reports the values of the dependence measure in sample industries. Table 6 in the Data Appendix offers the detailed descriptions of all measures used in this paper. It is worth to outline the interpretation of the coefficient of interest, β. Roughly speaking, the difference-in-differences estimator in the specification (10) splits the sample into four groups according to the magnitude of the fall in prices of information technologies and the dependence on these technologies. For each year, these four groups are composed of the industry-country pairs with high fall in prices and high dependence (HF&HD), industry-country pairs with high fall in prices and low dependence (HF&LD), pairs with 6 I show that the results are robust to excluding low wage occupations and computing industry-level employment as the sum of employment in high and medium wage occupations in the Online Appendix -Further Robustness Checks and Results. 7 I average the price of information technologies across sample countries and illustrate its trend over time in Figure 5 in the Online Appendix - Tables and Figures. low fall in prices and high dependence (LF&HD), and pairs with low fall in prices and low dependence (LF&LD). The coefficient β, in this respect, represents the difference in the trends of employment in occupation groups between HF&HD industry-country pairs relative to HF&LD industry-country pairs and LF&HD pairs relative to LF&LD pairs.
It is positive (negative) for an occupation group if employment in that group grows at a higher (lower) rate in HF&HD industry-country pairs relative to HF&LD industrycountry pairs than in LF&HD pairs relative to LF&LD pairs. I take the residuals from a regression of the share of employment in an occupation group on country-industry and country-year dummies to illustrate the existence of such differential trends. Panels A and B of Figure 1 show that there are such disparities in employment trends for high and medium wage occupations. These panels show that employment has increased (declined) more rapidly in high (medium) wage occupations in industries with high IT dependence relative to industries with low IT dependence, with the fall in IT prices. 8 Panel C of Figure 1 shows that there are almost no apparent differential trends in low wage occupations. Moreover, there seem to be no trends at all for low wage occupations, which suggests that, on average, employment in low wage occupations is likely to be not affected by the fall in IT prices, at least directly. 9

Results
Panel A of Table 4 presents the results from the estimation of the specification (10)  This suggests that, on average, information technologies are not likely to have direct effects on the share of employment in low wage occupations.
One way to compute the magnitude of these results is as follows. I take the countries and years where IT Price is the lowest and the highest and compute the difference between the levels of the inverse of IT Price for them. Further, I take the industries that rank the highest and the lowest in terms of the level of dependence on IT and compute the difference between dependence levels. Finally, I computê where ∆ stands for the difference operator. I also estimate the specification (10) for the shares of employment in high, medium and low wage occupations within each gender and age group. The results are reported in Panels B-F of Table 4. They are broadly consistent with the results for the shares of employment in occupation groups in Panel A, with some notable differences. The fall in IT prices has increased (reduced) the share of employment in high (medium) wage occupations among women by about 50 percent more than among men according to Panels B and C of Table 4. It has increased (reduced) the share of employment in high (medium) wage occupations among old workers by about 50 percent less than among young and among medium-age workers, according to Panels D-F . These differences are economically large. They are also statistically significant for genders and for age groups in medium wage occupations at least at the 10% level according to the standard t-test.
The differences are at the border line of statistical significance for age groups in high wage occupations.
A possible and common interpretation of these results is that the comparative advantage of performing tasks in medium wage and high wage occupation groups varies with gender and age. For example, males tend to be better endowed with hard motor skills (brawn) than females and these skills are commonly more important in many of the medium wage occupations. Meanwhile, women are argued to have an advantage in communication and social skills, which seem to be more important in high wage occupations. All else equal, the adoption and use of information technologies would then reduce employment in medium wage occupations and increase employment in high wage occupations among males less than among females. In turn, information technologies will have a lower effect on the share of employment in high and medium wage occupations among old if workers accumulate routine-skills more than other types of skills as they age (i.e., if α L,T R increases with age). 10

Further Results and Robustness Checks
The demand for IT is a potential source of reverse causality if industries' employment of different tasks affect it. Country-industry and country-year fixed effects in the specification (10) are likely to alleviate such reverse causality concerns given the sources of variation in IT prices. Nevertheless, I attempt to further circumvent the reverse causality concerns in two ways. Industries with the heaviest use of information technologies are the plausible candidates that affect prices of information technologies. In Panel A of  Table 4.
Omitted variables can be another source of endogeneity. It could be that the effects that the estimates in Panel A of Table 4 identify are not because of the fall in IT prices, but rather because of changes in the prices of new physical capital goods (see, e.g., Krusell, Ohanian, Ríos-Rull, and Violante, 2000). The movements in the prices of new physical capital goods could be the result of an ongoing investment specific technological change and could cause changes in the demand for physical capital and employment.
To test this hypothesis, I construct an industry-level measure for dependence on physical capital (net of information technology capital) and a country-year-level measure for the price of physical capital. The data for these measures are from the EU KLEMS database, and these measures are constructed in much similar way as the measures of dependence on IT (see Table 6 in the Data Appendix for details). I add to the specification (10) an interaction term between the measure of dependence of industries on physical capital and the inverse of the price of physical capital. Panel C of Table 5 reports the results. The coefficient on the interaction term between IT Dependence and 1/IT Price is very close to the coefficient reported in Panel A of Table 4. The coefficient on the interaction term between dependence on physical capital and the price of physical capital is significant for the shares of employment in high and medium wage occupations and has the same sign as the coefficient on the main interaction term. These results suggest that more ubiquitous processes, such as changes in the prices of physical capital, affect the share of employment in high and medium wage occupations and this is over and above the effects of information technologies. 11 According to the theoretical model, the parameter which measures the relative importance of routine tasks in value added, α T R , should be important for the analysis similarly to α IT . It can be easily shown from equation (5) that the fall in IT prices increases (reduces) the demand for abstract (routine) tasks more in industries with a low α T R than in industries with a high α T R . A supplementary test of whether the fall in IT prices has affected employment in high, medium and low wage occupations utilizes this variation.
The proxy for α T R can be constructed similarly to the proxy for α IT . Ideally, I need data for wages in routine intensive occupations in order to construct such a proxy. However, there are no data for wages in the ELFS database.
I use wage compensation of medium education-level (skill) employees obtained from the EU KLEMS database as a proxy for the wages in routine intensive occupations. According to Michaels et al. (2014), this can be a valid proxy because routine intensive occupations are the middle wage occupations and medium skill/education-level employment tends to be over-represented in these occupations at least in the US. 12 I measure α T R using the share of compensation of medium education-level employees out of value added in US industries, averaged over the period of 1993-2007. I add the interaction of this share with the inverse of IT Price to the specification (10). Panel D of Table 5 reports the results from this exercise. The results for the main interaction term are close to the main results in each regression. In turn, as expected, the estimated coefficient on this additional interaction term has a sign opposite to the sign of the main interaction term.
In the specification (10), country-year fixed effects will not fully capture the trends in relative wage rates if these vary among industries. Such a variation can be expected to be weak according to Kambourov and Manovskii (2009). However, it can confound the identification of β if the interaction term is correlated with it. In order to alleviate these concerns, I include in the specification (10) industry group-year dummies and present the results in Panel E of Table 5. I also add to the specification (10) the shares of wage and structural changes in the economy (e.g., Bárány and Siegel, 2018). 12 The Online Appendix -Education Levels within Occupation Groups offers an analysis of the relation between employment in middle wage occupations and medium education-level employment for European countries.
compensation of medium-and low-skill employees out of total wage compensation. The data for these variables are from the EU KLEMS database. Panel F of Table 5 reports the results. In both cases, the results are very similar to the main results suggesting that this is not likely to be a major concern.
The Online Appendix -Further Robustness Checks and Results provides a range of additional robustness check exercises and results. I also perform all of these robustness checks for the shares of employment in high, medium and low wage occupations among gender and age groups. For brevity, these robustness check results are not reported.

Conclusions
I use evidence from 10 Western European countries and the assignment of occupations into high, medium and low wage groups by Goos et al. (2014) and find that the share of employment in high wage occupations has increased with the fall in IT prices. The share of employment in middle wage occupations has declined with the fall in IT prices. In turn, I find no systematic evidence that the fall in IT prices affects the share of employment in the lowest paid occupations and that similar results hold within gender and age groups.
These results corroborate the polarization hypothesis.
I find certain important differences in gender and age groups, however. The fall in IT prices has increased (reduced) the share of employment in high (medium) wage occupations among males by about 50 percent less than among females. It has increased (reduced) the share of employment in high (medium) wage occupations among old workers by about 50 percent less than among young and medium-age workers. A possible common explanation for such results is that the comparative advantage of performing tasks specific to medium and high wage occupations varies with gender and age. All in all, these results suggest a need for a more nuanced view on the labor market effects of recent technological changes.    Table 6 in the Data Appendix for complete descriptions and sources of variables.

Tables and Figures
(1) (1) (1) (1) (2) (1) (1) Note: This table offers the results from the estimation of the specification (10) for the shares of employment in high, medium and low wage occupations in sample industries. In Panel A, dependent variables are the shares of employment. In Panels B and C, dependent variables are the shares of employment in high, medium and low wage occupations within males and females, respectively. In Panels D, E, and F , dependent variables are the shares of employment in high, medium and low wage occupations within young, medium-age, and old workers (in employment hours), respectively. See Table 6 in the Data Appendix for complete descriptions and sources of variables. All regressions include country-industry and country-year dummies and use the least squares estimation method. Standard errors are in parentheses. Standard errors are bootstrapped and two-way clustered at industry-and country-year-level. R2 (Partial) is the R-squared of the model where country-industry and country-year dummies have been partialled out. *** indicates significance at the 1% level, ** at the 5% level and * at the 10% level (1) (1) (1) (1) (2) (1) (1)

HF&HD -HF&LD LF&HD -LF&LD
Note: This figure illustrates the differences in the trends of employment shares of high, medium and low wage occupations in industry-country pairs with high and low fall in IT prices and high and low IT dependence. The curves with square tick symbols are the difference between the employment shares in industries with high IT Dependence and industries with low IT Dependence among countries and years where and when the fall in IT Price is relatively high (HF&HD -HF&LD). The curves with triangle tick symbols are the difference between the employment shares in industries with high IT Dependence and industries with low IT Dependence among countries and years where and when the fall in IT Price is relatively low (LF&HD -LF&LD). The employment shares in this figure are the residuals from an OLS regression of employment shares on country-industry and country-year dummies. In each of the four groups, these shares are averaged over countries and industries. An industry has high (low) dependence on IT if its IT Dependence is above (below) the median IT Dependence across industries. For a given year, the fall in IT Price in a country-year pair is relatively high (low) if the fall in IT Price (relative to its previous level) in that pair is lower (higher) than the median change in IT Price across countries in that year. It is sufficient to compare to the change because IT Price has declined everywhere. See Table 6 in the Data Appendix for complete descriptions and sources of variables and for the assignment of occupations into high, medium and low wage groups.

B Online Appendix -Further Robustness Checks and Results
This section presents the results from further robustness check exercises. It also offers additional results. I conduct robustness checks with respect to the sample of years, industries, and countries, identifying assumptions and variation in the data, and measures.  Table 7. Import competition can matter for employment shares in high, medium and low wage occupations (Autor, Dorn, and Hanson, 2015). According to equation (2), industry-level variation in factor input levels can affect the values of IT Dependence. This can bias the coefficient on the interaction term in the specification (10) in a non-trivial manner. In order to alleviate potential concerns, I make use of the fact that it is sufficient to have a correct ranking of industries according to their dependence on IT. Chen et al. (2016) and Kochanova (2016, 2017) utilize Left-hand side variables in all regressions are shares and are between 0 and 1. I estimate the specification (10) using Tobit with (0, 1) censoring. Panel A of Table 9 summarizes the results. These are almost identical to the main results.
I also estimate the specification (10) for all NACE 1-digit industries for which there are data in the ELFS and EU KLEMS databases. Panel B of Table 9 reports the results, which are very close to the main results.

B.A Not Reported Robustness Checks and Results
I continue performing robustness checks, but do not report the results for brevity. Offshoring can matter for employment shares according to Goos et al. (2014), although its effect in industries with different levels of dependence on IT is not a priori clear. To check whether offshoring can affect the results, I exclude from the analysis occupations which have offshorability score higher than the 75th percentile of offshorability index offered by Goos et al. (2014). Excluding these occupations does not have any significant effect on the results.
In the main text, I compute employment shares in industry-occupation pairs using (usual) weekly hours worked. I have checked that the results are robust to using the number of persons employed instead of the number of hours.
The possible changes of employment in manual tasks are disregarded in equation (6). I use all the available temporal variation in order to identify β. This allows to fully utilize the significant and omnipresent advances in information technologies over the sample period. I also attempt to estimate β using long differences. I take the differences between sample initial and end values in industry-country pairs in the specification (10) and estimate it for those differences including industry-fixed effects in it. This exercise provides estimates of β which have the same sign as the main estimates, but are not statistically significant. A reason for this is that taking long differences reduces the number of observations by about 10 times.
I perform all these robustness checks for the shares of employment in high, medium and low wage occupations within gender and age groups. I obtain results which are similar to the main findings.
There have been country-level changes in employment composition of genders and age groups in high, medium and low wage occupations. Such compositional changes can confound the results if they are more pronounced in industries with higher α IT and are not because of the fall in IT Price, but are correlated with it. To rule out such an explanation, I include an interaction of Industry Dependence and the country-level share of employment in the corresponding occupation group among gender and age groups in the specification (10). I obtain results which are qualitatively similar to the main findings.
Trends in female labor supply can confound the results for that group and the identified differences between genders. This concern is somewhat alleviated with the test showing that the results are robust to the inclusion of lagged dependent variable in the regression. I also obtain estimates of β, which are very similar to the estimates in Panel C of Table 4, when I estimate the specification (10) using data for the year 2000 and after. In this period, trends in female labor supply are much less pronounced. Cerina et al. (2017) show that polarization trends in female labor supply are more pronounced among married females than among not married females. I check that the estimates of β are very close the estimates in Panel C of Table 4, when I estimate the specification (10) for married, as well as not married, females.
There can be differences in intensive and extensive margins of adjustment of employment among males and females. The EU LFS database is a repeated cross-section, and one way to check this tests whether the results are different from the main results in Panels B and C of Table 4 among young males and among young females (age from 15 to 30). I find that the results for young males and among young females are not different from the main results.

Education-level Groups
I retrieve information from the ELFS database on the levels of education to check that the results for within education-level groups are similar to the main results in Panel A of Table   4. There are three levels of education in this database: pre-primary to lower-secondary (low; ISCED-97 0-2), secondary to post-secondary and non-tertiary (medium; ISCED-97 3-4) and tertiary (high; ISCED-97 5-6). I compute the number of hours worked in each of these education-level groups for all occupation group-industry cells in sample countries and years and the share of employment in high, medium and low wage occupations within these groups.
I estimate the specification (10) for the shares of employment in high, medium and low wage occupations within education-level groups. Similarly to the main results, the fall in IT prices has increased (reduced) the share of employment in high (medium) wage occupations and has had no robust effect on the share of employment in low wage occupations within these groups. Moreover, the effect of the fall in IT prices on employment shares of medium-educated workers in high and medium wage occupations is about twice as much as on employment shares among highly educated and low-educated workers. Michaels et al. (2014) use US data to argue that medium-educated (medium-skill) workers tend to be the most specialized in routine intensive tasks. Their argument sug-gests that medium-educated workers have the highest α T R . Admittedly, it might seem somewhat surprising then that the fall in IT prices has had a larger effect on mediumeducated workers from the lens of the theoretical model presented in the main text. This can warrant a further investigation. It suggests that either the theoretical model might not be very well suited for explaining differences in the results for low-, medium-and highly educated workers or the correspondence between education levels and occupation/task groups is weaker in European countries. The next section provides suggestive evidence for the latter.
The results from all not reported robustness checks are available upon request.
C Online Appendix -Education Levels within Occupation Groups Michaels et al. (2014) offer evidence from 11 developed countries (including Western European countries) that information and communication technologies (ICT) have changed skill demand. According to their findings, these technologies have increased the demand for highly educated workers at the expense of the demand for medium-educated workers. Michaels et al. (2014) argue that their results stem from a firm correspondence between education levels and employment in high, medium and low wage occupations groups and tasks performed in these occupation groups. In particular, they use US data to offer evidence that the fraction of highly educated workers is the highest within high wage occupations and, similarly, the fraction of medium-and low-educated workers is the highest within medium and low wage occupations, correspondingly.
This evidence motivates my use of the share of wage compensation of mediumeducated workers out of value added in US industries as a proxy for α T R . It also motivates my use of the wage compensations of medium-and low-educated workers as proxies for the compensations of workers in medium and low wage occupations. Nevertheless, I show below that the latter two proxies might be taken with some caution. The firm correspondence argued by Michaels et al. (2014) is not straightforward to replicate for occupations groups, industry aggregates, and the sample of European countries, which I use.   Table 15 reports correlations between wage rates of highly, medium and low educated workers and employment shares in high, medium and low wage occupations. The correlation between high-skill wage rate and employment share in high wage occupations is positive. Similarly, the correlations between medium-skill and low-skill wage rates and employment shares in medium and low wage occupations are positive. Taken together, this evidence suggests that there is a correspondence between education levels and wage/occupation groups in European countries. However, it is no not very strong.  Table 6 in the Data Appendix for complete descriptions and sources of variables. All regressions include country-industry and country-year dummies and and use the least squares estimation method. Standard errors are in parentheses. Standard errors are bootstrapped and two-way clustered at industry-and country-year-level. R2 (Partial) is the R-squared of the model where country-industry and country-year dummies have been partialled out. *** indicates significance at the 1% level, ** at the 5% level, and * at the 10% level.

D Online Appendix -Tables and Figures
(1) (2) (1)  Table 6 in the Data Appendix and Table 16 for complete descriptions and sources of variables. All regressions include country-industry and country-year dummies and use the least squares estimation method. Standard errors are in parentheses. Standard errors are bootstrapped and two-way clustered at industry-and country-year-level. R2 (Partial) is the R-squared of the model where country-industry and country-year dummies have been partialled out. *** indicates significance at the 1% level, ** at the 5% level, and * at the 10% level.  Table 6 in the Data Appendix and Table 16 for complete descriptions and sources of variables. All regressions include country-industry and country-year dummies. Standard errors are in parentheses. Standard errors are clustered at industry-level in Panel A. The estimation method is least squares in Panel B, and standard errors are bootstrapped and two-way clustered at industry-and country-year-level. R2 (Partial) is the R-squared of the model where country-industry and country-year dummies have been partialled out. *** indicates significance at the 1% level, ** at the 5% level, and * at the 10% level.    Table 6 in the Data Appendix and Table  16 for complete descriptions and sources of variables.

HF&HD -HF&LD LF&HD -LF&LD
Note: This figure illustrates the differences in the trends of employment shares of high, medium and low wage occupations within genders in industry-country pairs with high and low fall in IT prices and high and low IT dependence. The curves with square tick symbols are the difference between the employment shares within genders in industries with high IT Dependence and industries with low IT Dependence among countries and years where and when the fall in IT Price is relatively high (HF&HD -HF&LD). The curves with triangle tick symbols are the difference between the employment shares within genders in industries with high IT Dependence and industries with low IT Dependence among countries and years where and when the fall in IT Price is relatively low (LF&HD -LF&LD). The employment shares in this figure are the residuals from an OLS regression of employment shares on country-industry and country-year dummies. In each of the four groups, these shares are averaged over countries and industries. An industry has high (low) dependence on IT if its IT Dependence is above (below) the median IT Dependence across industries. For a given year, the fall in IT Price in a country-year pair is relatively high (low) if the fall in IT Price (relative to its previous level) in that pair is lower (higher) than the median change in IT Price across countries in that year. It is sufficient to compare to the change because IT Price has declined everywhere. See Table 6 in the Data Appendix for complete descriptions and sources of variables. See Table  6 in the Data Appendix for the assignment of occupations into high, medium and low wage groups.

HF&HD -HF&LD LF&HD -LF&LD
Note: This figure illustrates the differences in the trends of employment shares of high, medium and low wage occupations within age groups in industry-country pairs with high and low fall in IT prices and high and low IT dependence. The curves with square tick symbols are the difference between the employment shares within age groups in industries with high IT Dependence and industries with low IT Dependence among countries and years where and when the fall in IT Price is relatively high (HF&HD -HF&LD). The curves with triangle tick symbols are the difference between the employment shares within age groups in industries with high IT Dependence and industries with low IT Dependence among countries and years where and when the fall in IT Price is relatively low (LF&HD -LF&LD). The employment shares in this figure are the residuals from an OLS regression of employment shares on country-industry and country-year dummies. In each of the four groups, these shares are averaged over countries and industries. An industry has high (low) dependence on IT if its IT Dependence is above (below) the median IT Dependence across industries. For a given year, the fall in IT Price in a country-year pair is relatively high (low) if the fall in IT Price (relative to its previous level) in that pair is lower (higher) than the median change in IT Price across countries in that year. It is sufficient to compare to the change because IT Price has declined everywhere. See Table 6 in the Data Appendix for complete descriptions and sources of variables. See Table  6 in the Data Appendix for the assignment of occupations into high, medium and low wage groups. Note: This figure illustrates the trends in the shares of employment in high, medium and low wage occupation groups. These employment shares are averaged over the sample countries. See Table 6 in the Data Appendix for the assignment of occupations into high, medium and low wage groups. Note: This figure illustrates the evolution of the price of information technologies relative to the price of value added p IT (IT Price). This relative price is averaged across countries. See Table 6 in the Data Appendix for complete descriptions and sources of variables.

IT Price Ind
The price of investments in information technologies relative to the price of value added in sample industries. In contrast to IT Price, this variable is not averaged across industries. Source: EU KLEMS.

IT Investments
The ratio of real investments in information technologies and real value added in sample industries. Source: EU KLEMS.