Unraveling wage inequality: tangible and intangible assets, globalization and labor market regulations

We are witnessing an increasingly intense debate centered around the impact of artificial intelligence, automation technologies and robotics on economic growth, inequality and society as a whole. Economists, analysts, journalists and policymakers are split on the consequences of the introduction of these new technologies, with both optimists and pessimists. The former argue that, in the next decades, we will see a boost in productivity and new job opportunities, most of which are currently hard to envisage (e.g., Brynjolfsson and McAfee 2014; Baldwin 2019), while the latter predict significant job destruction and a sharp increase in income inequality (e.g., Freeman 2015; Frey and Osborne 2017; Berg et al. 2018).

While efforts have been made on the theoretical front to understand the mechanisms through which new technologies are shaping the functioning of modern labor markets (see, for instance, Acemoglu and Restrepo 2019, 2018; Nakamura and Zeira 2023), the empirical evidence is far from conclusive. For instance, by focusing on robots, quantitative studies (such as Graetz and Michaels 2018; Acemoglu and Restrepo 2020) capture only the ‘tangible’ part of technical change (Nelson and Phelps 1966), with employment, productivity and wages being affected as a result of investments ‘embodied’ in new machinery.

Moreover, interest has arisen in assessing the contribution of specific forms of investments previously not well acknowledged and measured, i.e., intangible assets (e.g., McGrattan and Prescott 2010, 2014). Consequently, researchers have questioned whether the labor-market effects of ‘intangible’ technical change (e.g., Corrado et al. 2009; Haskel and Westlake 2018), such as software and R &D, affect workers in a similar manner to tangible investments or not. In this respect, Blanas et al. (2020) provide evidence in support of the hypothesis that software displaced medium- and low-skilled workers, whereas Michaels et al. (2014) identify a negative impact of R &D on the share of the wage bill captured by medium-skilled labor, suggesting a polarizing effect of R &D. Furthermore, as stressed by Haskel and Westlake (2018), the impact of intangibles is expected to lead to a rising premium for well-educated workers, insofar as specific education and skills are required for managing these new technologies. The longstanding controversy about ‘disembodied’ and ‘capital-embodied’ technical change continues to be relevant in these discussions, with questions remaining over whether technological progress is independent of (i.e., Hicks-neutral) or tied to specific capital inputs in the economy and the consequences of this for labor market outcomes (e.g., Greenwood et al. 1997; Hercowitz 1998). Overall, the effect of technological progress may depend on how the two types of technical change—i.e., broken down into tangible and intangible capital assets—affect different kinds of workers, by either enhancing or limiting their relative importance in the production process.

A further potentially important driver of labor market outcomes addressed in a separate literature, focuses on the role of trade and particularly participation in Global Value Chains (GVCs). In this regard, as pointed out by Van Reenen (2011), trade with low-wage countries could force firms in advanced economies to ‘innovate or die’—producing, among others, significant impacts on the skill structure of labor demand, wages, inequality and productivity (see, among, others, Wood 1995; Hijzen et al. 2005; Hijzen 2007; Foster-McGregor et al. 2013; Michaels et al. 2014; Lopez Gonzales 2015).

Relatedly, the contributions of labor market institutions suggest that these have significant effects on living standards, productivity and social cohesion, especially in European economies (e.g., Koeniger et al. 2007; Betcherman 2012). In addition, several studies have highlighted how labor market institutions, when expressed, for example, through increased unionization, provision of minimum wages, balanced working hours and well-structured active labor policies, can favor a reduction in wage and income disparities (e.g., Jaumotte and Buitron 2015; Salverda and Checchi 2015).

The above discussion highlights that technological progress develops in various directions, with distinctions between tangible and intangible, as well as embodied and disembodied discussed in the literature. These different forms of technology potentially impact upon labor markets, and on different types of workers in labor markets, in varied ways. Beyond these broad categories, more specific technologies, such as robots and artificial intelligence (AI), are further expected to impact upon labor markets differently, with Webb (2020) arguing that while robots impact upon lower-skilled tasks, AI is targeted at high-skilled tasks. Despite such observations, much of the existing literature focuses on only a subset of technologies or bundles various different technologies together when considering technology’s impact on labor markets. This paper builds upon this existing literature by decomposing technology along different dimensions, examining the relationship between these different technologies and labor market outcomes. Specifically, we examine the relationship between relative wages and the following dimensions of technology: ICT Capital; Software and databases; R%D capital stocks; and Robots. Through this analysis, our aim is to identify and understand the relative effects of these distinct technologies on the relative wages of different types of labor, providing a more nuanced comprehension of the complex relationship between different types of capital and labor market outcomes. Moreover, when considering the relationship between these different dimensions of technology and labor market outcomes, we account for the other major drivers of labor market outcomes considered in the literature, i.e. labor market institutions and trade, further focusing on the role of GVCs in the context of trade. The approach thus builds upon the technology-skill complementarity and polarization of the labor force frameworks (e.g., Autor et al. 2006; Goos and Manning 2007; Goldin and Katz 2009; Goos et al. 2009; Acemoglu and Autor 2011), but is more comprehensive in considering and decomposing technologies that may affect labor markets. The paper thus contributes to various strands of the literature: from the “capital-embodied” technical change to technology-skill complementarity and polarization of the labor force frameworks, as well as work examining the impact of automation and digital technologies, institutions and international trade on labor market outcomes. The paper further has important policy relevance, with concerns increasing around the impacts of these diverse technologies. By improving our understanding of how distinct types of technology impact upon different types of labor, the study provides insights into the range of policy responses that may be needed to tackle the increased investments in these different technologies.

In terms of data, the research relies on a panel of 17 mostly advanced European economies and 5 industries, using annual data over the years 2008–2017. In performing the empirical investigation, we exploit the EU KLEMS (2019) database, which explicitly groups fixed capital stocks into tangible and intangible assets, according to Haskel and Westlake (2018). Additionally, by integrating data on operational stocks of industrial robots from the International Federation of Robotics (IFR), we have the opportunity to detect the influence of advances in robotics, ICT, R &D and Software & Databases (S &DB) as different, independent proxies for tangible and intangible technologies, respectively. To disentangle the effects of the above-mentioned drivers on relative wages, we adopt a two-step estimation strategy. Firstly, we simultaneously estimate a system of wage premium equations by making use of seemingly unrelated regressions (SUR) to deal with correlations in the error terms across equations. The second step involves an instrumental variable approach to address potential endogeneity concerns, using Japanese robot density as an instrument for robot adoption in the sample of European countries included in the analysis.

The empirical analysis reveals that decomposing technology and considering the independent effects of different dimensions of technology is important in understanding their overall effects on labor market outcomes. In fact, the multiple aspects of technology are found to show varied relationships with relative wages. Robots, for instance, are associated with an increase in relative wages for high-skilled labor, but also with a compressed wage gap between medium- and low-skilled workers, suggesting a polarization effect. R &D also seems to benefit higher skilled labor disproportionately. Conversely, ICT is associated with a lower wage premium for high-skilled labor, while S &DB is associated with lower relative wages for low-skilled labor. The results thus underscore the importance of decomposing technology, with the average association of technology on relative wages hiding a great deal of variation across technology types.

The paper is structured as follows: Section 2 reviews the relevant literature; Sect. 3 describes the data employed in the analysis; Sect. 4 illustrates the theoretical and empirical frameworks, as well as the estimation strategy; Sect. 5 presents the baseline findings; Sects. 6 and 7 deal with the results based on the instrumental variable approach; finally, Sect. 8 provides a set of policy recommendations and concludes.

The empirical literature investigating the role of (automation) technology, globalization and institutions in affecting labor market outcomes constitutes a large and growing body of research. Starting from the influential contributions by Solow (1960) and Jorgenson (1966), the controversy of ‘embodied’ versus ‘disembodied’ technical change revolved around whether technological progress affects productivity and economic growth through changes in physical capital goods (embodied) or independently of capital (disembodied). On the one hand, embodied technical change suggests that technological advancements are tied to specific capital investments, such as machinery or equipment. This perspective implies that capital accumulation is necessary for productivity growth. On the other hand, disembodied technical change posits that technology can improve productivity independently of capital investments. According to this view, technological progress can lead to efficiency gains and productivity improvements even without significant changes in the capital stock (e.g., Greenwood et al. 1997; Hercowitz 1998; Boucekkine et al. 2003).

Meanwhile, since the seminal work by Griliches (1969) on capital-skill complementarity, many scholars have examined the potentially biased effects of technology on the relative demand for skilled workers. In particular, the evidence in the early nineties provided by Katz and Murphy (1992) and Bound and Johnson (1992) gave a new momentum to considering the efficacy of the skill-biased technical change hypothesis in explaining the observed rising trend in wage inequality across countries and within groups (for exhaustive surveys on this subject, see Chusseau et al. 2008; Acemoglu and Autor 2011). More recently, an alternative to the skill-biased technical change hypothesis has been proposed that attempts to provide an explanation more suitable for the recent observation of declining relative demand and wages of middle-skilled workers—the so-called job polarization phenomenon—in developed countries in particular (e.g., Autor et al. 2003; Goos et al. 2009; Acemoglu and Autor 2011). The routine-biased technical change hypothesis (Autor et al. 2003) argues that recent technological change, including artificial intelligence, robots and ICT developments more generally, allows for the replacement of workers doing routine tasks, which are often tasks undertaken by middle-skilled workers.

Among the proxies for automation technologies employed in the empirical literature on the skill composition of labor demand and wages, the focus has mostly been placed on computerization, ICT, R &D expenditure and patents (e.g., Berman et al. 1994; Morrison Paul and Siegel 2001; Chennells and Van Reenen 2002; Michaels et al. 2014; Mann and Püttmann 2021). Relying upon new data from the International Federation of Robotics (IFR) on industrial robots, progress has been achieved in the study of the impact of this contemporary automation wave on labor market outcomes, albeit with mixed results. Pioneering works in this field are Acemoglu and Restrepo (2020) and Graetz and Michaels (2018), who find evidence, respectively, of negative effects of robotics on wages and employment in the US and a positive influence on labor productivity growth in a panel of 17 countries. With specific reference to European economies, the findings are even less clear-cut. For instance, Chiacchio et al. (2018) point out a significant employment reduction as a result of increasing robot density (measured as the number of robots per thousand workers), an effect that is felt most strongly by middle-educated workers. By contrast, Dauth et al. (2021), analyzing 402 German local labor markets over the years 1994–2014, observe no effect of industrial robots on total employment, but adjustments in the composition of aggregate employment—specifically, job losses in manufacturing are offset by gains in the service sector. Similarly, by using data on employment from the European Labour Force Survey, Klenert et al. (2023) find that the adoption of an additional robot is associated, on average, with the employment of five additional workers.

Although the effects of robotics on labor market outcomes have been analyzed from various angles (for recent surveys, see Mondolo 2022; Yan and Grossman 2023), the empirical literature on the effects of robots on wage inequality still appears to be scarce. In this respect, Barth et al. (2020) point out that robotization increases the wages of high-skilled workers relative to low-skilled workers within Norwegian firms. Likewise, Aksoy et al. (2021) highlight how industrial robots are associated with a significant increase in the gender pay gap across 20 European economies.

Most of the technologies so far discussed, such as computerization, ICT and robots are tangible in nature, but since the contribution of Corrado et al. (2005), a new emphasis has been placed on the incidence of so-called ‘intangible’ investments, previously not appropriately classified and counted by business and national accounts. As argued by Haskel and Westlake (2018), intangibles are characterized by unique economic properties, among which are their complementarity, especially with well-educated and high-paid workers, as well as their tendency to generate knowledge and/or idea spillovers among firms and to trigger a “competitive process of investments in continuous product improvement” (Haskel and Westlake 2018, p. 41). These features could help explain a variety of economic phenomena such as economic growth, secular stagnation, and rising income and wealth inequality (e.g., Corrado et al. 2009; Glaeser 2011; Bessen 2016; Song et al. 2019). In particular, relying on a panel of 10 developed countries and 30 industries over the period 1982–2005, Blanas et al. (2020) find that software, as a proxy for intangible technology, is associated with an increase in the demand for high-skilled workers only, while the tangible component of ICT has a positive impact on the demand for all workers types. However, to the best of our knowledge, nothing has been done in the empirical literature to understand whether intangibles may be strongly complementary (substitutable) to high-skilled (low-skilled) workers, thus exacerbating wage inequalities, or be associated with a polarization of the wage distribution, and whether the tangible component of ICT, by contrast, may negatively affect wage dispersion (e.g., Acemoglu and Restrepo 2020, 2022).

In addition to technological advances, the many dimensions of globalization are thought to play an important role in affecting wage and income disparities (for recent reviews of the literature see, for instance, Kurokawa 2014; Nolan et al. 2019). According to the traditional Heckscher–Ohlin–Samuelson (HOS) model, trade openness is expected to benefit the abundant factor, which in developed countries would tend to suggest a rise in demand for, and therefore the return to, skilled relative to unskilled labor. In this respect, Wood (1995) analyzed the labor market effects of north–south trade, providing evidence of a significant impact of trade in reducing low-skilled employment in manufacturing in the North. Other studies have tended to provide confirmatory evidence of an effect of trade openness and/or liberalization on the skill-premium in developed countries, although the effects tend to be smaller than those found for technology. For instance, Harrigan and Balaban (1999) observe that capital accumulation and the decline in traded goods prices increased the earnings of well-educated workers in the USA, while Robbins (1996) and Beyer et al. (1999) highlighted a growth in the skill-premium in Chile. More recently, Michaels et al. (2014) and Epifani and Gancia (2008) find results suggesting polarizing and skill-biased effects of international trade, respectively. Goos et al. (2014) also find evidence to suggest that offshoring can lead to job polarization.

In reconsidering the traditional HOS trade-based approach, which has attracted considerable criticism (e.g., Berman et al. 1998; Goldberg and Pavcnik 2007), attempts have been made to provide new explanations for the role played by different forms of trade engagement—in particular, international outsourcing and offshoring—in driving wage inequality worldwide (for a survey, see Hummels et al. 2018). As argued by Feenstra and Hanson (1996), developing economies have played an increasing role in producing more skill-intensive inputs as a result of outsourcing by advanced economies, generating a rise in the relative demand for skilled workers and the skill-premium in both developed and developing countries. Conversely, Grossman and Rossi-Hansberg (2008) offer a different explanation: by assuming that the prices of goods remain unchanged, a cost decrease in offshoring produces an increase in unskilled activities offshored to developing countries. This, in turn, causes a rise in profits and sector expansion for those industries that heavily employ unskilled labor, pushing up the sector’s demand, productivity and wages, while leaving demand unchanged in industries relying on skilled labor. Therefore, through this channel, the skill-premium decreases.¹

The evolution of the new patterns of globalization has been embodied by Gereffi and Korzeniewicz (1994) in the concept of GVCs. According to Amador and Di Mauro (2015), GVCs describe “the full range of activities undertaken to bring a product or service from its conception to its end use and how these activities are distributed over geographic space and across international borders”. The role of geographically dispersed production stimulated many studies to assess the impact of GVC participation on earnings and wages (e.g., Baumgarten et al. 2013; Hummels et al. 2014; Parteka and Wolszczak-Derlacz 2015). In a recent contribution, Wang et al. (2021) develop a model suggesting that GVC participation is associated with higher profitability, which in turn leads to demands for higher wages (based upon a fair wage assumption). Given the higher bargaining power of skilled workers, GVC participation would increase the skill-premium. However, little has been done to quantify the effects of GVC participation on inequality. For instance, Lopez Gonzales (2015) measure backward GVC participation using the foreign value-added share of a country’s gross exports and find that increased GVC participation is associated with a narrowing wage gap between skilled and unskilled labor in both developed and emerging economies—a finding in line with the theoretical predictions by Grossman and Rossi-Hansberg (2008). Conversely, as pointed out by Timmer et al. (2014), the expansion of emerging economies resulted in an increased global supply of low-skilled workers, consequently exerting downward pressure on the relative wages of less-skilled labor in developed countries.

Finally, a relevant role in affecting labor market outcomes is played by institutions and regulations (for a survey of studies on the effects of labor market institutions on living standards, productivity and social cohesion, see Betcherman 2012), such as measures of employment protection legislation (EPL). Existing studies find mixed evidence on the effect of labor market protection on labor market outcomes. A number of studies have demonstrated a significant and substantial effect of strong labor market protections in mitigating wage differentials, as shown, for example, by Koeniger et al. (2007) for a sample of 11 OECD countries over the period 1973–1998. Conversely, using data for a sample of 20 OECD countries spanning the years 1973–2011 and indicators for regular and temporary contracts, Sparrman and Rossvoll (2015) find that the two indicators of labor market restrictions have opposite impacts on wage inequality, with EPL for temporary contracts shrinking the wage gaps and EPL for regular contracts intensifying them.

In summary, existing work on the relationship between technology and labor markets, often focuses on one particular dimension of technology (e.g., robots) or uses a broad definition (e.g., patents), thus neglecting the various types of technology and how they may impact upon labor markets in different ways. Moreover, the review of the literature suggests that there is relatively little evidence linking the various types of technology to relative wages specifically. In what follows, therefore, we build upon and extend existing work linking technology to labor markets, empirically examining the relationship between relative wages and different aspects of both tangible and intangible technologies, as well as different globalization measures and labor market regulations on determining wage differentials dynamics.

The empirical analysis relies on annual panel data for 17 mostly developed European economies and 5 industries—classified according to the one-digit-level NACE Rev. 2 (ISIC Rev. 4)—spanning the period 2008–2017.²

Data are collected and integrated from various sources. The main dataset is the EU KLEMS (2019) database, which provides information on skill composition, employment, labor compensation, hours worked, real fixed capital assets and value added by country-industry-year. The EU KLEMS dataset combines information on the shares of labor compensation and hours worked for three different worker types, which are distinguished on the basis of their educational attainment: university graduates; secondary and post-secondary education; and primary and lower secondary education.³ Such a decomposition allows for a multifaceted investigation of the dynamics of skill-premia, analyzing whether workers are affected differently by tangible and intangible technologies, as well as by globalization and labor market regulations.

Relative wages are calculated as the ratio of the higher to the lower-educated hourly wage, along the three dimensions (i.e., high- to medium-skilled workers, \(w_h/w_m\); high- to low-skilled workers, \(w_h/w_l\); and medium- to low-skilled workers, \(w_m/w_l\) ).⁴ Relative skill supplies (i.e., the quantity effect), included in the models as controls, are measured by the ratios of hours worked in each analyzed category—namely the ratio of high-skilled to medium-skilled hours worked (H/M), the ratio of high-skilled to low-skilled hours worked (H/L), and the ratio of medium-skilled to low-skilled hours worked (M/L). It is important to note that the data we have are limited to information on employment by education level rather than by specific skill sets or tasks that workers undertake. While this serves as a valuable proxy for examining labor aspects discussed in recent task-based literature, it needs to be kept in mind that this mapping from education to tasks and activities is not perfect.

As for capital inputs, based on Haskel and Westlake (2018), the EU KLEMS database groups asset types into tangibles and intangibles. Specifically, the tangible category includes ICT net of Software & Databases (i.e., hardware) and non-ICT (comprising, among others, transport equipment and total non-residential investments) capital stocks. The intangible assets contain S &DB and R &D capital stocks.⁵ All capital intensity variables are taken as a ratio to total hours worked (in millions of hours).⁶

The second source of data is the International Federation of Robotics (IFR) for the stock of industrial robots by country-industry-year. IFR data are broken down by industrial branches and classified according to ISIC Rev. 4, which makes them highly compatible with EU KLEMS. However, due to limitations in the number of industries covered, the merger with EU KLEMS is possible only for the sectors reported in Table 7 of the Appendix.⁷ In the benchmark analysis, the robot density variable (R) is computed as the number of industrial robots⁸ per million hours worked, rather than the number of persons engaged, on the grounds that workers in different countries/industries may vary in the quantity of hours worked (Graetz and Michaels 2018). As observed by Blanas et al. (2020) and Jungmittag et al. (2019), robots are widely deployed in heavy industries, as a form of automation that links machinery (non-ICT capital) and software. Nonetheless, because of the tangible nature of robots, the inclusion of robot density in the analysis is aimed at isolating potential independent effects on the skill-premia and, following Graetz and Michaels (2018) and de Vries et al. (2020), assessing the robot high-skill complementarity or “hollowing-out” of middle-skilled workers hypotheses.

In the second stage of our analysis, we examine the role played by globalization, and in particular overall trade openness and participation in GVCs, in strengthening or mitigating the wage premia. For this purpose, we use data from the Trade in Value Added (TiVA) database (OECD, 2022) to measure the extent of trade openness at the country-industry level. By aggregating information at the one-digit level, the overall measure of international trade (GLOB) is calculated as the sum of intermediate imports and total (i.e., intermediate plus final) gross exports expressed as a share of real gross value added. According to Epifani and Gancia (2008) and Michaels et al. (2014), we would expect either high-skill-oriented or a polarization of the wage distribution, as a result of higher trade openness. Thus, whether GLOB affects the skill-premia positively or negatively, for the three dimensions of wage inequality, is an empirical question.

From the same source we gather information to account for participation in GVCs. Specifically, we collect data on domestic value added embodied in foreign final demand (\(FFD\_DVA\)) and foreign value added embodied in domestic final demand (\(DFD\_FVA\)). These two indicators can be interpreted, respectively, as “exports of value-added” and “imports of value added”—capturing forward and backward linkages in GVCs, respectively. The (total) GVCs participation indicator is computed as the sum of \(FFD\_DVA\) and \(DFDA\_FVA\) expressed as a ratio to real gross value added. The inclusion of GVC variables, in the third stage of the analysis, has the goal of detecting whether and how a diverse form of engagement in trade (i.e., other than GLOB) impacts upon different kinds of workers.

Finally, from the OECD we employ data on Employment Protection Legislation (EPL) in the third stage of the study, where the impact of labor market institutions on the skill-premia is assessed. Borrowing from IMF (2016) and Hantzsche et al. (2018), and making use of information from the Eurostat Labour Force Survey (EU-LFS, 2020) database, we construct two sector-specific indices of EPL for permanent and temporary workers, \(EPL\_PERM\) and \(EPL\_TEMP\). By including the sector-specific measures for EPL in the models, we test the hypothesis that the recent findings of a negative relationship between EPL and skill-premia (e.g., Koeniger et al. 2007) are confirmed when the extent of labor market regulations are proxied by two separate, sector-specific indicators, one for each group of workers. As an additional check, we also consider an overall sectoral EPL index (\(SECT\_EPL\)), given by the sum of \(EPL\_PERM\) and \(EPL\_TEMP\).

Real price variables are expressed in PPP adjusted 2005 international dollars, with the PPP conversion factors taken from Inklaar and Timmer (2014). The benchmark sample consists of 751 observations. Summary statistics, by country and industry, for the levels of the variables included in the empirical analysis are reported in Tables 8, 9, 10 and 11 of the Appendix, respectively.

To set the underlying rationale of our empirical investigation, we rely on the contributions of Acemoglu and Autor (2011) and Blanas et al. (2020), proposing a simple theoretical framework analyzing the effects of both tangible (including robots) and intangible assets on relative wages.

Let us assume that output per-worker, y, is obtained by an intensive production function, f, combining multiple factors¹²:where \({\textbf {A}}\) is a vector of efficiency parameters, \({\textbf {K}} \, = \{ R,S\) &DB, ICT, R &\(D,N\_ICT \}\) indicates a vector of capital per-worker inputs and h, m and l denote high-, medium- and low-skilled labor inputs, respectively.

We consider that labor markets are competitive and inelastically supply h, m and l to firms. Each skill group is required to perform a set of (non-mismatching) tasks, \(z \; = (z_h,z_m,z_l)\), and is paid its marginal product (i.e., \(w_h\), \(w_m\) and \(w_l\)). We also assume that initially there is no automation (\({\textbf {K}} ={\textbf {0}}\)) that can replace for labor in performing tasks. First-order conditions of profit-maximizing and price-taking firms imply that the relative wages can be expressed as a function, g, of relative supplies and allocation of tasks to skill groups in which these exhibit a comparative advantage. For instance, the skill-premium between high- and medium-skilled workers is given by:where \(\theta _{hm} \, = \frac{h}{m}\) represents the relative supply of high- to medium-skilled workers, while \(z_h\) and \(z_m\) denote tasks allocated to h and m, respectively. The relative wages of medium- to low-skilled workers and high- to low-skilled workers, \(w_m/w_l\) and \(w_h/w_l\), are obtained analogously as in Eq. (2).

Once we allow for automation replacing tasks (\({\textbf {K}} \ne {\textbf {0}}\)) previously performed by labor, according to the job and wage polarization framework proposed by Autor et al. (2003), medium-skilled workers—typically employed in intermediate routine or codifiable tasks—should be the group affected the most by the current wave of innovations, such as robots and digital (intangible) technologies. As a result, we can first expect that \(w_h/w_m\) increases and \(w_m/w_l\) decreases. As for the \(w_h/w_l\), the direction of the impact of technological progress is an empirical question, depending on whether m are closer substitutes for h or l.

Additionally, some specific types of technological progress may reveal peculiar, strong complementarities or substitutability with a particular skill group. The extent of these will depend on the relative importance of the specific channels through which capital impacts on different types of labor. These channels include a productivity effects, which operates via both labor-augmenting and capital-augmenting technological change, a substitution effect involving certain tasks of distinct labor types being undertaken by different forms of capital, and an effect linked to the emergence of new tasks associated with specific forms of capital that can increase the demand for particular types of labor (Acemoglu and Restrepo 2019). In light of this, the polarization of the wage distribution, as in the scenario of robots replacing middle-skilled workers, is only one possible outcome. Such an eventuality thus motivates our intention of dealing with a variety of tangible and intangible technologies in a comprehensive setting, with the objective of discovering their potential differential impact on wage premia dynamics along the three possible dimensions.

To this end, and building upon the outlined theoretical framework and empirical works of, among others, Goldin and Katz (2009), Michaels et al. (2014), Glitz and Wissmann (2021), and Graetz and Michaels (2018), the estimated (system of) three equations (in logarithmic form) accounting for the evolution of skill-premia can be explicitly formulated as follows:where \(c = 1,...,C\), \(s = 1,...,S\) and \(t = 1,...,T\), indicate, respectively, country, industry and time; \(\alpha _{n,c}\), \(\beta _{n,s}\) and \(\gamma _{n,t}\) (with \(n = \left\{ 1,2,3\right\} \)) are country, industry and time fixed effects, respectively, to control for cross-country and cross-industry unobserved heterogeneity, and to capture time varying unobserved factors, such as global shocks. The dependent variable is the skill-premium, \(w_i/w_j\) (with \(i = \left\{ h,m\right\} \) and \(j = \left\{ m,l \right\} \))—namely \(w_h/w_m\), \(w_h/w_l\) and \(w_m/w_l\); \({\textbf {K}}\) is a vector of the main explanatory variables, including R¹³ and the EU KLEMS capital intensity variables, i.e., ICT, R &D and S &DB, with \({\delta }\) being the vector of coefficients of interest; \({\textbf {X}}\) denotes the vector of control variables, including non-ICT capital intensity (\(N\_ICT\))¹⁴ and relative supplies of labor (h/m; h/l; m/l); and \(\epsilon _n\) are well behaved error terms (with \(n = \left\{ 1,2,3 \right\} \)).

The three skill-premium relationships in Eq. (3) are both individually and simultaneously estimated using Ordinary Least Squares (OLS) and SUR techniques (Zellner 1962), respectively. The latter may prove to be more reliable and efficient than OLS, for example, if all relations are influenced by common factors, which can induce a correlation of the error terms across the equations. According to Goldin and Katz (2009) and Glitz and Wissmann (2021), identification of the model given by Eq. (3) relies on the assumption that the relative skill supplies are inelastic in the short-run (i.e., predetermined), stemming from past investment decisions in education and training. Therefore, under such an assumption, the wage premia and relative skill supplies are not jointly determined.¹⁵ In Sect. 6, we further discuss additional endogeneity concerns, such as those related to productivity shocks—reflected in wages—that may impact upon technology adoption. From a cross-industry perspective, however, these issues should be less concerning, due to the lower substitutability and/or mobility of inputs across sectors (e.g., Battisti et al. 2022). In any case, we do not claim any causal effect at this stage, as our intent is to highlight the statistical link between the different indicators of capital and technology and relative wages.

This section presents and discusses the baseline estimates of the study, where the focus is placed upon the role played by tangible and intangible technologies. Table 1 reports the OLS and SUR results, in Panels I–II and Panel III, respectively, for the three wage premium equations described in the previous section: high- to medium-skilled workers (\(w_h/w_m\)), in Column (a); high- to low-skilled workers (\(w_h/w_l\)), in Column (b); medium- to low-skilled workers (\(w_m/w_l\)), in Column (c). Initially, in Panels I and II of Table 1, we present the results of the specification in Eq. (3), estimated by OLS, and evaluate whether the outcomes are sensitive to the inclusion of additional control variables in the models (i.e., non-ICT capital intensity and relative supplies), whereas in Panel III, the SUR estimation allows us to control for correlation of the disturbances across equations. In this respect, the Breusch–Pagan test strongly rejects the null hypothesis of contemporaneously independent disturbances across the equations—providing support for the adoption of SUR estimates against OLS.

Turning to our main variables of interest, in all cases the estimated coefficients on the technology-skill complementarity effect of robot density (R)—our first measure of tangible technologies—are in line with our expectations and suggest that they widen the skill-premia of high-skilled with respect to both medium- and low-skilled workers (see Columns (a) and (b)), results in line with those of Graetz and Michaels (2018), Blanas et al. (2020) and Dauth et al. (2021). Such an outcome may reflect a complementary relationship between robot density (R) and high-skilled workers or a substitution effect with respect to low- and medium-skilled workers. In elasticity terms, all else being equal, a 1% increase in R is linked, on average, to a growing wage gap of 0.08% and 0.03%, for \(w_h/w_m\) and \(w_h/w_l\), respectively. Moreover, as shown in Column (c) of Table 1, we detect a negative correlation between robot density and \(w_m/w_l\). In particular, a 1% increase in robot density is linked, on average, to a declining \(w_m/w_l\) of approximately 0.05%. This finding, combined with the coefficients presented in Column (a), suggests that robots are associated with a polarization of the wage distribution, disproportionately affecting medium-skilled workers. Such results are in line with both predictions and empirical findings by Acemoglu and Autor (2011), Acemoglu and Restrepo (2018) and de Vries et al. (2020), who highlight a severe decline in the employment share of routine manual task-intensive jobs.

Focusing on the role of R &D capital—the first proxy for intangible technologies—we observe coefficients in Columns (a) and (c) of Table 1 consistent with a polarizing effect. This is shown by the positive coefficient of R &D on \(w_h/w_m\) and the negative coefficient on \(w_m/w_l\). At the same time, the evidence is somewhat weak with the coefficient on \(w_m/w_l\) being insignificant and that on \(w_h/w_m\) being only marginally significant.

Results on the intangible part of ICT—i.e., S &DB—present a consistent pattern across the different specification and indicate that intangible ICT is associated with falling relative wages for low-skilled workers. This is true with respect to both high-skilled and medium-skilled wages, with the coefficients being consistently significant. In this case, ceteris paribus, a 1% increase in the intensity of S &DB use is associated, on average, with an increase in \(w_h/w_l\) and \(w_m/w_l\), on average across panels, of around 0.035% and 0.025%, respectively.

On the contrary, the tangible component of ICT capital—i.e., hardware—appears to improve, to a greater extent, the wages of lower-educated workers, relative to high-skilled workers, as shown in Columns (a) and (b) of Table 1. Such results may be due to the role of ICT as a general-purpose technology,¹⁶ as well the fact that such technology has reached maturity and become pervasive across European economies, facilitating the adaptability and skill upgrading opportunities even for middle- and low-skilled workers. In turn, this relative enhancement in productivity may translate into an improvement in wage differentials vis-à-vis high-skilled labor (e.g., Aghion and Commander 1999; Conceição and Galbraith 2012; Acemoglu and Restrepo 2020, 2018).

Overall, our findings suggest that considering relative wages along three dimensions and disentangling the roles played by different kinds of technological advances in a systematic and comprehensive analytical framework, can favor a better understanding of the dynamics of wage dispersion within the labor market induced by tangible and intangible automation technologies. In particular, the evidence so far indicates that different technologies do not affect wage differentials in the same way, being characterized by specific complementarities and/or substitutabilities with respect to different types of workers.¹⁷

However, several factors can lead to biased results when using OLS and SUR to estimate Eq. (3). The subsequent section highlights some of these issues and explores an alternative instrumental variable (IV) approach to deal with some of these factors.

In our framework, endogeneity bias could arise due to reverse (or bidirectional) causality, with relative wages potentially influencing the direction of technological progress. If there exists a bidirectional causal relationship, OLS/SUR can result in biased estimates. For example, technological advancements, including heightened automation, may lead to increased productivity and higher (relative) wages for skilled workers. These, in turn, might incentivize firms to invest in new skilled labor technologies, which may (or may not) impact on the wages of low- and medium-skilled workers.¹⁸ Additionally, omitted variables, such as market competition, government policies, tax incentives or financial market conditions, can affect both technology adoption and wage differentials, introducing a bias in the estimates. Hence, reverse causality and omitted variable bias pose a risk to the credibility of OLS/SUR-based estimates in establishing a proper causal relationship (i.e., from technology to relative wages).

To mitigate the aforementioned concerns, we employ a two-stage least squares (2SLS) strategy. The first-stage equation can be formulated as follows:where \({\textbf {Z}}\) indicates the vector of other capital intensity variables (i.e., ICT, R &D and S &DB), fixed effects and controls outlined in Eq. (3), while \(R_J\) denotes robot penetration in Japan in sector s and year t, which is used as an IV for R.

Subsequently, the second-stage equation can be expressed as:where \(\hat{R_J}\) represents the predicted variable from Eq. (4) and the other explanatory variables remain the same as in Eq. (3).

\(R_J\) can be considered a reliable IV for R only if the exclusion restriction holds, i.e., automation in Japan is likely to influence European relative wages only through the impact on its robotization process. However, this assumption deserves more scrutiny. In general, robotization in Japan may prove to be a suitable instrumental variable for robotization in Europe for several reasons. First, this country is known for its advanced robotics industry and is a global leader in industrial robots production (e.g., Nof 1999; Gasparetto and Scalera 2019). Further, according to IFR, 2019, Japan is more advanced in robotic technologies compared to our sampled economies. Second, Japan has a long history of adopting and implementing industrial automation, with a high density of robots used in manufacturing processes (Mandfield 1989). Therefore, this suggests that the use of robots is an important factor driving productivity and efficiency improvements in this country, which can be applicable to European industries undergoing robotization. Third, Japan is also the largest exporter of industrial robots in the world, providing 45% of the global supply IFR (2022). Such a market influence implies that the adoption of industrial robots in Europe may be affected by trends and advancements originating from Japan.

Building upon these considerations, using robot density in Japan (\(R_J\)) as an instrumental variable for their European counterparts (R) can provide a relevant and appropriate approximation of the impact of robotization within these industries. In terms of excludability, Japanese robot density is not directly related to the European skill-premia. In other words, this implies that the effect of industrial robots in Japan on European labor market outcomes (including wage differentials) operates only through their influence on the level of robotization. Quantitatively, we expect \(R_J\) to be positively associated with R in the first-stage equation (i.e., Eq. (4)), reflecting the technological catch-up process in automation pursued by the European countries in our sample. Therefore, by using Japanese robot density, we isolate the exogenous source of variation originating from a country at the forefront of robotization (a similar approach is adopted by Acemoglu and Restrepo 2020, who make use of robot penetration trends in five advanced European countries ahead of the USA in robotics.).

Results from 2SLS estimates of the three skill-premium relationships in Eqs. (4)-(5) are reported in Table 2.¹⁹ The table includes two sets of results, both excluding (Panel I) and including (Panel II) additional control variables in the model. In both cases, as expected, the first-stage coefficient on Japanese robot density (\(R_J\)) is positive and strongly significant. In terms of magnitude, this implies that, ceteris paribus, a 1% increase in \(R_J\) is associated with an increment in the range of 0.34\(-\)0.43% in European robot density (R).²⁰ Moreover, the F-test statistic from the first-stage regression provides strong evidence against the hypothesis of weak instruments.

The second-stage estimates substantially confirm the core findings, as discussed in the previous section, with the exception of the coefficient associated with R in Column (b), which loses its statistical significance for the relationship involving \(w_h/w_l\). Furthermore, the 2SLS regressions reveal that the coefficients estimated through OLS/SUR were upward (downward) biased for the relationship between R and \(w_h/w_m\) (\(w_m/w_l\)). In fact, upon comparing the outcomes of Columns (a) and (c) of Table 2 (Panel II) with the counterparts of Table 1 (Panels II and III), the 2SLS estimated coefficients suggest that a 1% increase in R is associated, on average, with an increment in \(w_h/w_m\) of about 0.065% and a decline in \(w_m/w_l\) of approximately 0.07%. In other words, medium-skilled workers experienced similar relative losses compared to both high- and low-skilled labor.

By and large, the outcomes of our IV regressions allow us, on the one hand, to mitigate endogeneity problems characterizing the relationship involving robotization and relative wages and, on the other hand, substantiate the hypothesis that investments in tangibles and intangibles technologies are likely to produce a variety of effects on the three dimensions of wage inequality. Particularly, we observe that not all types of technology considered impact relative wages in the same way. For example, in line with the predictions of Haskel and Westlake (2018), intangibles have a detrimental effect (although the evidence is rather weak in the case of R &D) on the relative wages of lower-skilled workers, while ‘tangible’ automation—as proxied by robots—is associated with a polarization of the wage distribution, as conjectured, inter alia, by Acemoglu and Autor (2011).

In this section, we consider extensions, which also serve as robustness checks, to the benchmark analysis described above. Specifically, the first extension, reported in Sect. 7.1 takes into account the cost of robots in our study. The second extension, presented in 7.2, involves the inclusion of globalization, while Sect. 7.3 further incorporates variables capturing sector-specific proxies for labor market regulations. Through this analysis we are interested in examining whether the inclusion of these additional drivers of the skill-premia impact—in terms of robust correlations—upon the estimated relationships between the different technologies and wage premia described in the previous section.

The growing concerns about the issues of artificial intelligence, robotics, automation and digital innovation on the future of people’s working lives, supplemented by the well-known puzzling influence of global trade and offshoring, has recently led many researchers to question and investigate the real effectiveness and magnitude of the impact exerted by these powerful economic forces within the labor market, especially in developed countries. The results of several studies have strengthened such concerns, leading to calls for policies directed at protecting jobs and industries from new and/or foreign threats. By contrast, other scholars reject such a pessimistic view, claiming that many of the fears are clearly unfounded.

In this paper, we contribute to the ongoing debate by studying the effects of automation technologies, as well as different forms of international trade engagement and labor market institutions, on the wage premia, relying on the technology-skill complementarity and polarization approaches. The empirical analysis is performed using annual data for a panel of 17 mostly advanced European economies and 5 industries over the period 2008–2017. According to the recent literature, new technologies are split into tangibles and intangibles, globalization is considered through the overall trade openness and GVC participation channels, while the impact of labor market institutions on wage disparities is evaluated by making use of sector-specific measures of Employment Protection Legislation. In order to detect potential specific effects of the main determinants of wage gaps for different worker types, we break down relative wages in three categories (high- to medium-skilled, high- to low-skilled and medium- to low-skilled labor) and simultaneously estimate a system of equations employing SUR techniques to take into account correlation of the error terms across equations. However, to mitigate endogeneity concerns, an instrumental variable approach is mainly used throughout the empirical investigation.

The core results of our analysis can be summarized as follows. First, intangible technologies, as proxied by Software & Databases and R &D capital intensity, exhibit low-skill substitutability and a polarization of the wage distribution, respectively. Second, we find that robotization is associated with a polarization of the wage distribution, producing heterogeneous effects on the relative wage distributions as well. Third, through channels such as “job insecurity” and “the fear of automation,” which could potentially alter the bargaining power of employers, the decline in robot prices exerts additional pressure on the relative wages of medium-skilled workers. Fourth, the role of globalization on the dynamics of the wage differentials depends upon the specific measure considered. Greater trade openness appears to be characterized by both high- and medium-skill-biased patterns, while higher participation in GVCs predominantly harms lower-skilled labor. Finally, employment protection rules prove to be effective in mitigating medium- and low-skilled wage differentials.

From a policy perspective, the main challenge is represented by the effects of intangible technologies and robotization. As our findings suggest, high-skilled workers typically benefit from technological progress (consistent with the extensive existing empirical literature on this topic). This implies, in terms of policy implications, that the strong substitutability we observe between low-skilled workers and Software & Databases (and, albeit to a lesser extent, between medium-skilled labor and R &D) related process innovations, as well as the “hollowing effect” of robotization call for the adoption of institutional measures and actions aimed at investing in education and skills training for less-skilled workers, particularly given that intangible technologies and robots are likely to pervade the workplace even more in the future. Overall, policymakers will need to play a crucial role in ensuring that the economic benefits stemming from new technologies will not be focused on a small elite and further research should be devoted to understanding the exact mechanisms by which rising automation might lead to new job opportunities or destruction.

Furthermore, our findings point to harmful effects of trade openness and GVC participation upon lower-skilled labor. Blanchard and Willmann (2016) suggest that subsidizing human capital investments and/or providing temporary wage top-ups for these particular categories of workers may be a relevant policy.

In essence, our empirical investigation indicates that the influence of tangible and intangible technologies can either be positive or negative in affecting the dynamics of the skill-premia, with the effects depending on the specific dimensions, characteristics and economic mechanisms underlying them. Such a conclusion suggests that there may exist a third way, which lies between the technological optimists and pessimists, whereby the different dimensions of technology (and globalization) affect workers in varied ways.