1 Introduction

Since the start of the 2008 financial crisis, “European countries have variously faced economic stagnation, rising inequalities, worsening social conditions, and strains on public services” (della Porta 2021, p. 325), exacerbated by a polarization process resulting in the spatial concentration of economic development, wealth, and population (Görmar et al. 2019). In terms of macro areas, we are witnessing a concentration of employment, skills, competitiveness, and industrial specialization in a “German-centered core” whereas the “Southern periphery” is lagging behind (Cirillo and Guarascio 2015). However, diverging trajectories also arise at a lower scale. After a period of relative convergence until the end of the 1980s, EU regional inequalities have importantly increased in the new millennium (Storper 2018; Fratesi and Perucca 2019), resulting in a complex patchwork of spatial patterns: income and employment differ not only between states and regions but also between core and peripheral areas within the same region, as well as between dynamic metropolitan regions and stagnating ones (Iammarino et al. 2019).

These growing inequalities are at the core of an intense debate (Piketty 2013; Alvaredo et al. 2017; Stiglitz 2013), which identifies among their main causes the impetuous technological development that has occurred over the last 40 years (Autor et al. 2003; Hanson and Harrison 1995; Jovanovic 2009; Van Reenen 2011), of which robotization represents one of the most notable steps. As a matter of fact, the recent technological development has produced such structural changes as to upset the patterns of development of modern economies. Within this debate, the robotization of the production process has assumed relevance, renewing the discussion between techno-optimists (Caselli et al. 2021; Ballestar et al. 2020; Graetz and Michaels 2018; Dauth et al. 2019; Brynjolfsson and McAfee 2014; Mokyr 2010; Jovanovic and Rousseau 2005) and techno-pessimists (Fernald 2014; Manyika et al. 2017; Gordon 2012; Arntz et al. 2016; Frey and Osborne 2017), as Andrews et al. (2016) suggest.

Technological change, in fact, causes structural change, which directly affects regional economies (Gräbner et al. 2020) and, as a result, indirectly translates into variations of the aggregate and sectoral demand for goods and services. Accordingly, Compagnucci et al. (2022) empirically show that the structural composition of regional economies has been one of the main causes of the asymmetric effects of the 2008 financial crisis. In this regard, what makes the difference is the capacity of a region to lead and shape the technological change or, on the contrary, to passively suffer its effects. In other words, the impact of robotization on employment and income will depend on the regional capacity to maximize the complementarities between humans and robots, while minimizing the substitution effect. Obviously, the level of human capital and of specific economic activities (i.e., high-tech manufacturing and knowledge intensive services), which are unevenly distributed across regions, plays a crucial role in this sense. Still, the presence of these factors is importantly affected by regional policy concerning R&D and education levels.

The recent rise of the smart specialization strategy within the EU regional and urban policy reforms (McCann and Ortega-Argilés 2016) is a case in point, since it aims at strengthening research, technological development, and innovation at the regional level. The same applies to the commonly known "Industry 4.0" strategy, which encourages industrial sectors to adopt the new technologies to move towards higher-value digitized products and processes, supported by regional and local authoritiesFootnote 1. Given the level of knowledge and skills that this type of specialization requires, education policies take on particular importance, although a different mix of competencies between local and national governments in the EU countriesFootnote 2. For instance, in Italy, secondary school is a matter of regional authorities, whose decision-making process is based on local peculiarities and needs (Bruno et al. 2022). This policy thus has consequences in terms of accessibility, not only to secondary but also to tertiary education, as well as in coping with the possible mismatch between the skills provided by the education system and those required by the regional labor market. By doing so, it can help create the conditions for the complementarity effect between human beings and machines. On the contrary, in case of their poor design due to budgetary constraints or lock-in in public decision-making, the substitution effect could prevail with negative outcomes in terms of citizens' welfare. These conclusions draw attention to the need for regional place-based policies in the EURO area, so that the robotization process does not turn out to be yet another shock harbinger of territorial inequalities (Aghion et al. 2019).

Despite the importance of robotisation in our societies, contributions on the topic tend to be anecdotal and focused on productivity and employment outcomes at the macro and national levels only, also due to the lack of data on robot at a regional scale (Green Leigh and Kraft 2017). This implies little attention to the regional level and related policies, which, in contrast, is the context where policy makers must not only respond to potential employment losses but also propose strategies to remain competitive in the changed contextual conditions.

Against this background, the aim of our work is to contribute to the debate by showing how the outcomes of robotization at the regional level are conditioned by long-term policy choices such as the investment in R&D and the capacity of the education system to raise the qualitative and quantitative level of human capital. Our analysis of the Euro Area regions in the last 20 years shows that there is a positive interaction between the above-mentioned policies and the robotization process of the different industrial sectors. While, in general, a higher exposure of European regions to robotization, computed according to Acemoglu and Restrepo (2020), has compressed households’ income against non-significant positive effects on employment, those regions that have invested more in R&D and human capital have been able to more than compensate for the income losses and to improve employment levels. These results support the idea that regional policies play a key role in influencing regional development trajectories, and that, if well designed, they can help mitigate the negative externalities generated by the robotization process.

The remainder of the paper is structured as follows. Section 2 reviews the relevant literature. Section 3 discusses the data employed and the empirical strategy. Section 4 discusses the main results, whereas Section 5 provides some final remarks.

2 R&D, education, and robotization

In 1969, Atikinson and Stiglitz, breaking with neoclassical orthodoxy, argued that improvements in technology could have affected the use (and hence the remuneration) of the different factors of production in a non-uniform manner, thus violating the assumption of technological neutrality proposed by Hicks (Acemoglu et al. 2015). Such predictions came true in the late 1980s, when an unprecedented growth in R&D investment and in the use of technology in the production process occurred (Kim and Nelson 2000; McConnell and Perez-Quiros 2000). These processes jointly caused a structural shift towards more innovation- and knowledge-oriented sectors in advanced countries (van Winden 2007; Raspe and Van Oort 2006; Florida 2002).

According to Acemoglu (1999), this radical change had a substantial impact on income distribution. Since technological progress is fundamentally based on competencies, it has mainly rewarded the most skilled and educated workers, while replacing with machines the less skilled and poorly educated, who perform repetitive work. Acemoglu and Loebbing (2022), by developing an assignment model of automation, show that automation occurs (regarding medium-skilled competencies) when: (1) high-skilled workers have a comparative advantage relative to capital in complex tasks; and (2) low-skilled workers’ compensations are lower than their productivity and the cost of capital for their replacement. If the cost of capital decreases (or its productivity increases), then we will experience more automation, pushing workers at the top and the bottom of the task distribution. This finally leads to unemployment, wage polarization, and a reduction in real wage of workers at risk of substitution. These theoretical considerations are empirically discussed by Acemoglu and Restrepo (2021), who argue that the changes that occurred in the US wage structure in the last 40 years depended between 50% and 70% on the relative wage decline of workers doing repetitive tasks. These theoretical considerations are discussed empirically by Acemoglu and Restrepo (2021), who argue that the changes in the wage structure in the United States over the past 40 years have been 50%-70% dependent on the relative wage decline of workers performing repetitive tasks in sectors which experienced a rapid automation.

Such results have important spatial implications. The endowment of physical and human capital, in fact, is not uniform across regions and factors’ mobility is far from the perfect one predicted by neoclassical models. Highly skilled and educated workers are concentrated in some of the largest urban and metropolitan regions of advanced countries (Krätke 2007; Florida 2002; Moretti 2012, 2013). Baum-Snow and Pavan (2012) show that the overall increase in the spatial divide of wages that occurred between 1979 and 2007 in the United States was significantly correlated with city size, which positively affected nominal wages, housing prices, and productivity. Kemeny et al. (2022), using the universe of patents between 1920 and 2010 in the US, found that disruptive innovation (as robotization is) is spatially clustered at its outbreak and plays a substantial role in regulating overall patterns of spatial output and income inequality. Thus, since technological progress is localized (Atikinson and Stiglitz 1969) and benefits from scale and agglomeration economies, it tends to polarize the distance between regions endowed with different levels of physical and human capital (Fleisher et al. 2010).

Recently, the processes of polarization within and between social groups and regions has often been associated with the process of automation (Brynjolfsson and McAfee 2011; Graetz and Michaels 2017; Manyika et al. 2017), particularly in the manufacturing sector, where industrial robots operate. By industrial robot (definition ISO 8373 - International Federation of Robotics 2020) we mean an automatically controlled, programmable, multi-purpose manipulator, mobile on three or more axes. Based on five different types of kinematics, the modern industrial robot is the extreme evolution of the mechanical spinning machine generally identified, together with the railway, as the symbol of the First Industrial Revolution. A robot, however, differs significantly from the mechanical spinning machine, and from all the production automation technologies that have followed since 1700, since it is reprogrammable and usable for different purposes.

Robots, in short, contain not only "capital", but also "work", since they are able to perform different functions almost (if not totally) independently. They represent, potentially, a new type of worker that requires hardware and software maintenance but at the same time allows the elimination of all the costs linked to human "weaknesses" (from the need for breaks, rest shifts, sick leave, and holidays, to salary claims), which have negative effects on productivity and profits. It is no coincidence that the COVID-19 pandemic has pushed towards the robotization and automation almost all types of industries (Lund et al. 2020). Robots are, indeed, potential substitutes for human workers. Undoubtedly, industrial robots allow to produce more (when not even to perform tasks otherwise impossible for humankind) in less time, with fewer errors and lower costs than human labor (Dauth et al. 2019; Graetz and Michaels 2018). However, the literature does not agree on the possible economic and social consequences of the adoption of such technology in all productive activities currently performed by humans (see Mondolo 2021 for a comprehensive discussion of the relevant literature).

The positivist view generally identifies six possible mechanisms through which the adoption of capital goods allows increasing productivity while maintaining constant (when not increasing) the demand for labor. Specifically, this can occur in conjunction with: (1) the creation of employment in the sector that produces robots meant to replace the human workforce (Say 1964); (2) the reduction of production costs in the sector that adopts the robots, which results in a reduction of the output price and, therefore, in the possible expansion of the sector itself (Hall and Hefferman 1985); (3) a situation of non-instantaneous competitive convergence generating extra profits and allowing capital accumulation which in turn can boost investments and, ultimately, positive employment effects (Van Reenen 1997); (4) the lowering of wages allowing the reabsorption of workers in other sectors (Pigou 1933); (5) the generation of positive externalities related to the increase in income, which in turn stems from increased productivity (see Piva and Vivarelli 2017a for a detailed discussion of these issues); (6) the introduction of new goods and services and therefore the creation of new business sectors (Schumpeter 1918).

However, as Freeman (2016) writes: "There is [...] nothing in economics that guarantees that the humans displaced from jobs by robots will end up with new jobs that pay as much as their former jobs [...]". In this regard, the literature has already shown how technological innovation can act as a dislocating force of labor when the expansion of the sectors involved is not fast enough. This can happen for different reasons: due to institutional constraints (Herwartz and Niebuhr 2011), or input shortages (Attinasi et al. 2021), but also in the presence of mature sectors and saturated markets unable to expand their output (Bessen 2019). Moreover, the negative effect of technological innovation on employment and income may prevail in the absence of income-oriented public support for unemployment, training of workers, and stimulation of demand for the new emerging sectors (Delli Gatti et al. 2012). As pointed out by Acemoglu and Restrepo (2020), the same technology may boost employment or can reduce the demand for labor according to the capacity of the industry and workers to move to non-automated tasks faster than the pace of automation.

Hence, even robotization corroborates the soundness of the theoretical shift from a neutral technological change towards a “biased and directed technological change” which, in addition, is importantly localized (Acemoglu et al. 2015, p. 445). It follows that, as Nuccio et al. note (2020), more than assessing the direct employment impact of robotization, it is worth considering the polarization effect “between economies that grow because of the early adoption of automation (and associated organizational change) and economies that shrink because they were not able to catch up” (p. 5). Regional systems, in conclusion, can facilitate or hinder the adoption of new technology and the rise of new sectors, and institutions can play an important role in such a context (Nuccio et al. 2020).

Focusing on policies, several macro-strands arise in the literature. The first one concerns the investment in R&D, which may produce effects on points (1) and (6) of the positivist view previously outlined. The second regards policies aiming at strengthening the comparative advantage of the labor force, such as the financing of the education and training systems, which allows a qualitative-quantitative improvement of human capital, promoting the complementarity between labor and machine (point 5 of the positivist view). Both of these types of policies generally suffer from underinvestment, the level of which also depends on the different public–private mix that characterizes each economic system. R&D expenditure, for instance, results from both public and private investments. Although it is aimed at a specific goal (increasing the innovation level at the micro and macro levels), it generates a series of positive externalities for the whole community. However, as it is well established in the literature since Arrow (1972), private firms, in order to maximize their profit, usually invest in R&D less than the optimal value for the social planner. It follows that, where R&D expenditure primarily results from the private sector, the beneficial externalities for the community could be under-sized. In these cases, a direct action by the policy makers could be necessary.

Other kinds of polices were proposed to deal with robotization. About robot taxation, for instance, Zhang (2019) finds that it unconditionally reduces wage inequality, both with or without the presence of a labor union in the automated sector, and when introducing a regulated wage rate in that sector. In addition, the collective ownership of robots, along with the introduction of a universal income (which is further useful to support the aggregate demand during transition phases), as well as the management of the working time (Cséfalvay 2019; Belloc et al. 2020) could also be effective in mitigating the negative effects of robotization. However, unlike education/training and R&D expenditure policies, cases regarding the adoption of these policies are very limited and even less are the available data. Therefore, these options are not considered here.

On the contrary, the planning and adoption of R&D and education policies are unanimously reputed to be crucial, resulting in being more widespread, and thus allowing spatial comparisons. Already in 2010, the European Commission (2010) pointed out that even in times of fiscal austerity, there was a need to invest in education, R&D, innovation, and ICT in order to maintain the capacity to create employment and ensure the welfare of European citizens. For R&D, given the under-investment in knowledge (as a percentage of GDP) of the EU average (0.8% lower than in the USA and 1.5% lower than in Japan), it was believed that reaching the threshold of 3% of GDP would have allowed an increase in employment of about 3.7 million jobs.

Although the 3% target has not been reachedFootnote 3, nevertheless some researchers have highlighted a statistically significant relationship between R&D expenditure and employment performance, even though, in most cases, net of the heterogeneous effects arising from different sectors. Indeed, technological change and firm innovation "in some parts of the economy may either stimulate or come at the expense of employment in others" (Herstad and Sandven 2019, p. 376). Piva and Vivarelli (2017b), for example, based on a firm-level database including the largest European R&D investors in the period 2002 to 2013, find a positive relationship between the two variables. This relationship, however, only holds for high- and medium-tech manufacturing. In line with these results and with reference to a representative sample of Spanish manufacturing firms between 2002 and 2013, Pellegrino et al. (2019) show how this positive relationship only concerns high-tech manufacturing. Bogliacino et al. (2012), analyzing data concerning 677 European firms in the period between 1990 and 2008, obtain very similar results, and thus positive and significant effects in terms of R&D-related employment in the high-tech manufacturing and services sectors only, with zero effects on medium-low-tech manufacturing. Ciarli et al. (2018), with reference to the English Travel to Work Areas, highlight two underlying dynamics. On the one hand, the increase in R&D investment in functional areas with more routine jobs had negative effects in sectors with low-skilled employment, especially in non-tradable sectors (as transport and construction) and in services, but, at the same time, exerted a positive effect on the self-employment of younger workers. On the other hand, in functional areas with a lower rate of routine work, a positive relationship emerges between R&D and the demand for high-skilled workers versus a reduction in demand for low-skilled workers. Similarly, Aldieri et al. (2021), analyzing the regional expenditure in Finland between 2000 and 2013, find positive effects only for the demand of high-skilled jobs, but negative effects for low-skilled jobs. If, on the one hand, such empirical evidence clearly corroborates the effort of EU policies in stimulating R&D, on the other hand it highlights the potential risks for low-tech manufacturing, which still plays an important role in Europe, requiring a coordination effort with education policies to address the issue of low-skilled workers, as well as the need to introduce innovation within low-tech manufacturing (Cardamone et al. 2018).

There is ample evidence in the literature on how spending on R&D, education, and training of workers creates positive societal externalities. Increased investment in R&D should, in principle, facilitate the transition towards the technological frontier (Lee 2009; Hall et al. 2013) by exploiting and expanding the margins of complementarity between humans and machines. This also allows to benefit directly from the production of the capital goods underlying the 4IR, or from the creation of new business sectors made possible by technological advancement. Furthermore, increased investment in R&D should raise the demand for a more highly educated and skilled workforce (Autor et al. 1998), thereby augmenting the positive externalities exerted by the knowledge economy (OECD 1996; Engelbrecht 1997) with consequent multiplier effects on employment and wages.

Nonetheless, the role of human capital in avoiding or mitigating the employment costs associated with the automation of production processes is a well-known issue. Lankisch et al. (2019) empirically show that automation is responsible for the wage decline of non-skilled workers and, potentially, of the skilled ones, and for the raise of the skill premium. Its capacity to affect income inequality stresses the importance of investing in higher education, which, in turn, by preventing the substitution effect, increases the share of population benefiting from the automation-driven economic growth.

The hypotheses of skill-biased technological change (Acemoglu and Autor 2011) and of routine-biased technological change (Autor et al. 2003) are intrinsically linked to the need to increase the qualitative and quantitative level of human capital through education and retraining of the workforce where necessary. Against the automation process, the acquisition of new skills and abilities would allow workers to fill earlier non-automated occupations and/or functions, similarly to what happened with the migration of the workforce from the primary sector to industry following the mechanization of agriculture (Delli Gatti et al. 2012). Acquiring new skills would allow workers to "migrate" more easily to activities that are complementary with those that are automated. Although the two strands of literature differ substantially in the specific role attributed to skills and abilities versus tasksFootnote 4, both give human capital a central role. The inability of economic systems to provide for the needs of continuing education and (re)training becomes, de facto, one of the causes of worker dislocation and increasing inequality (Acemoglu and Restrepo 2018).

3 Robotization in European regions

To create a reliable proxy for regional exposure to robotization, we calculate the metric Adjusted Penetration of Robots (APR) at the sectoral level, following the approach suggested by Acemoglu and Restrepo (2020). For this purpose, we use two databases: the International Federation of Robotics (IFR) and the STructural ANalysis databases (STAN OECD).

3.1 Sectoral adjusted penetration of robots (APR)

The first step of the analysis consists of providing a measure for the APR at industry (i) level, in a given country (Italy = IT, for example), from t0 to t1. Specifically, according to Acemoglu and Restrepo (2020), the following specification is proposed:

$${\overline{\mathrm{APR}}}_{\mathrm{i},\left({\mathrm{t}}_0,{\mathrm{t}}_1\right)}^{\mathrm{IT}}=\frac{{\mathrm{M}}_{\mathrm{i},{\mathrm{t}}_1}-{\mathrm{M}}_{\mathrm{i},{\mathrm{t}}_0}}{{\mathrm{L}}_{\mathrm{i},{\mathrm{t}}_0}}-{\mathrm{g}}_{\mathrm{i},\left({\mathrm{t}}_0,{\mathrm{t}}_1\right)}\frac{{\mathrm{M}}_{\mathrm{i},{\mathrm{t}}_0}}{{\mathrm{L}}_{\mathrm{i},{\mathrm{t}}_0}}$$
(1)

where, M is the stock of robots (IFR), L is the number of persons employed (per 100,000) (STAN), g is the growth rate of the output (STAN).

This measure of robots' penetration combines the industry-level variation in the usage of robots, baseline employment shares, and an adjustment for the overall expansion of each industry's output, given by the last term of the equation.

Acemoglu and Restrepo (2020) calculate APR at the US level and the average of five European countries (Denmark, Finland, France, Italy, and Sweden). In general, it is possible to compute this average for N countries:

$${\overline{\mathrm{APR}}}_{\mathrm{i},\left({\mathrm{t}}_0,{\mathrm{t}}_1\right)}^{\mathrm{N}\mathrm{countries}}=\frac{1}{\mathrm{N}}{\sum}_{\mathrm{j}\in \mathrm{N}}\left[\frac{{\mathrm{M}}_{\mathrm{i},{\mathrm{t}}_1}^{\mathrm{j}}-{\mathrm{M}}_{\mathrm{i},{\mathrm{t}}_0}^{\mathrm{j}}}{{\mathrm{L}}_{\mathrm{i},{\mathrm{t}}_0}^{\mathrm{j}}}-{\mathrm{g}}_{\mathrm{i},\left({\mathrm{t}}_0,{\mathrm{t}}_1\right)}^{\mathrm{j}}\frac{{\mathrm{M}}_{\mathrm{i},{\mathrm{t}}_0}^{\mathrm{j}}}{{\mathrm{L}}_{\mathrm{i},{\mathrm{t}}_0}^{\mathrm{j}}}\right]$$
(2)

In our analysis, we consider nine countries (Austria, Belgium, Finland, France, Germany, Greece, Italy, Netherlands, and Portugal) whose base value results from the average of their concerned industries (as defined by IFR) in t0. Table 1 reports the calculation of APR, namely the change in robot usage per 100,000 employees according to the different industries considered. Following Acemoglu and Restrepo (2020), and according to the fact that the effect of robotization may take some years to have an impact on economic variables and that APR yearly variation may be misleading, we undergo staked differences analysis. We therefore divide the period into four main sub-periods: 2000–2004, 2004–2008, 2008–2012, and 2012–2017. This allows us to have two periods before and two periods after the 2008 financial crisis. Moreover, according to Acemoglu and Restrepo (2020), we imputed “non-specified” robots from the IFR database to sectors according to their distribution in each country.

Table 1  APR per industry and period (Delta robots per 100.000 employed)
Full size table

Before describing the empirical strategy, it is necessary to discuss some preliminary peculiar values. The value 0 in "Chemicals and pharmaceutical products" is the result of a variation equal to 0 in the stock of robots in the first two periods under analysis. This result can hardly be explained by the literature on patents that identifies the pharmaceutical sector as one of the sectors more exposed to labor-saving innovations (Montobbio et al., 2022). It follows that either a large part of those patents does not translate into capital goods, or, more likely, that the adoption of robots in the sector has started only recently due to: (a) the complexity of the necessary authorizations, and (b) the fact that pharmaceutical production lines are generally dedicated to a single product, making the use of multifunctional robots less interesting for the production activityFootnote 5.

3.2 Regional exposure to robotization (EXP)

Following Acemoglu and Restrepo (2020), sectoral APR is combined with the employment shares of different sectors and regions to calculate a proxy for regional exposure to robotization (EXP):

$${\mathrm{EXP}}_r={\sum}_{\mathrm{i}\in \mathrm{I}}{l}_{r,i}\times {APR}_i$$
(3)

where lr,i are the employment shares of industry i in region r. Acemoglu and Restrepo (2020) define this measure of exposure to robotization as "constructed from the interaction between baseline industry shares in a local labor market and technological possibilities for the introduction of robots across industries" (p. 3).

To compute the regional employment shares for the sectors reported in Table 1, we first used the OECD Regional database to compute the shares of agriculture, construction, mining and utilities, services, and manufacturing, and then we reconstructed the shares of the different manufacturing sectors using Eurostat's Structural Business Statistic (SBS) database.

It is important to note that the first database (OECD regional) contains data for all employed workers (self-employed and employees) while the second (SBS) only reports data for employees. Despite this limitation, the resulting shares can be considered a good approximation of the sectoral composition of regional economies.

As the spatial classification between Eurostat and the OECD is not uniform due to the repeated spatial variations over the last 20 years, and to the fact that, in some cases, the OECD considers larger regional partitions than Eurostat, regions have been aggregated at the level of the OECD classification. Moreover, given the impossibility of obtaining available data at the NUTS 2 level for Ireland and Finland, we consider Ireland at the national level and only two regions for Finland. In addition, overseas regions were excluded from the analysis. Finally, economic sectors (particularly manufacturing sectors) were translated from NACE Rev. 1.1 to NACE Rev. 2 and then aggregated according to the nomenclature proposed by IFR. For more detail about the aggregation methods used for territorial and sectoral breakdowns, see Compagnucci et al. (2021).

This process allowed us to calculate the Regional Exposure to Robotization (EXP) of the EURO area regions using the number of robots installed in the 16 industrial sectors reported in Table 1. It shows that the only relevant sector for which no analysis could be provided is "Coke and refined petroleum products" due to the lack or limited availability of related employment data at the regional level.

As data on robotization at the regional level are not available, EXP should be considered as a proxy of the potential exposure to robotization, or, in other words, the theoretical risk to which workers are exposed, based on the sectoral composition of the regional economy and the installations of robots (considered at the European level) in different sectors.

The measures of EXP in the four periods are reported in Fig. 1. Again, it is useful to emphasize some stylized facts. First, it is worth remembering that the different sectors contribute to EXP according to the amplitude of the sectoral APR and the importance of the different economic sectors at the regional level. It follows that the most exposed regions are those more specialized in the production of "Rubber and plastic products" and "Transport equipment". The EXP, in fact, is greater in Germany, in northern Italy, in the Spanish Atlantic area, and in various regions of France. For the same reason, Basilicata is the only region in southern Italy to be highly exposed, given the presence of several companies related to the automotive industry.

Fig. 1
figure 1

Regional exposure to robotization (delta robots per 100,000 employed). Source: authors' original calculations on IFR, STAN and SBS data. N.B.: In this case, the last period has been weighted by 4/5 to take into account the longer time interval. N.B.2: In the maps "PT20" "PT30", and "ES70", i.e., the furthest island, have been removed for the sake of clarity. They are, however, included in the analysis

Full size image

Secondly, it should be noted that, after the 2008 financial crisis, there has been a surge in EXP in every region. According to the equations used for APR (1) and EXP (2), this is not a consequence of an increase in robot installations, but rather of a negative output growth rate.

4 Strategy of analysis and econometric models

The second step of the analysis consists of estimating different models, starting from the following generic panel specification:

$${\displaystyle \begin{array}{c}\Delta {Y}_i^{t,t-1}={\mu}_i+{\gamma}_t+{\beta}_0\cdot {EXP}_i^{t,t-1}+{\beta}_1\cdot \mathrm{INT}\_\mathrm{DUMMY}\cdot {EXP}_i^{t,t-1}+\sum_{c=1}^C{\theta}_c\cdot {Control}_{i;t-1}^c+\\ {}+\rho \cdot \sum_{j=1}^n{w}_{i,j}\cdot \Delta {Y}_j^{t,t-1}+{\nu}_{i,t}\\ {}{\nu}_{i,t}=\lambda \cdot \sum_{j=1}^n{w}_{i,j}\cdot {\nu}_{j,t}+{\epsilon}_{i,t}\end{array}}$$

where:

  • \(\Delta {Y}_i^{t,t-1}\)is the variation in the dependent variable which, depending on the case, coincides with:

    1. 1.

      the households’ income per-capita growth rate (constant prices)Footnote 6.

    2. 2.

      the variation in the employment rate.

    3. 3.

      the variation in the unemployment rate.

    4. 4.

      the variation in the labor force participation rate.

    5. 5.

      the labor utilization (worked hours/population) growth rate.

  • All the dependent variables are considered in four periods using a stacked differences panel analysis (hence t = 2000, 2004, 2008, 2012, 2017). As the last period is longer than the others, its changes (∆Y and EXP) are weighted by 4/5 (the same applies to Table 2).

  • μi e γt are the regional and time fixed effects, respectively.

  • EXP is our variable of interest, which interacts in the different models considered with two different dummies (INT_DUMMY) that distinguish between: (1) regions with a high level of investment in R&D (HIGH_R&D = 1) and the others (HIGH_R&D = 0); (2) regions with a highly educated workforce (HIGH_EDU = 1) and the others (HIGH_EDU = 0). In the case of R&D expenditure, we compute the yearly median for each region over the entire 2000–2017 period. Each of them is successively compared with the overall median. When the regional median is higher than the overall median, than we assign the value 1 to the relative dummy, otherwise the assigned value is 0. The same applies to the percentage of the population (between 25 and 64 years) with at least an upper secondary education level and the corresponding dummy related to the level of education. By estimating β0, we obtain the coefficient measuring the correlation of EXP with the dependent variable for the regions with dummy equal to 0, while β1 regards the difference between the coefficient of the regions with dummy equal to 0 and those with dummy equal to 1. The sum of the two coefficients explains the overall effect of robotization risk in the regions with dummy equal to 1.

  • Control are the regional controls. Specifically: GDP per-capita growth rate (which allows us to control for the economic cycle); employment shares in manufacturing, utilities and mining, agriculture, construction (which allows us to split the exposition to robotization per se from the mere industrial composition – services are the omitted reference sector); number of patents; population density; percentage of population over 65; male–female ratio. Except for the GDP growth rate, all variables’ values are considered at t0.

  • Finally, ρ is the spatial correlation of the dependent variable, and λ the spatial correlation of the errors. By estimating these parameters, we further check the robustness of the results with respect to potential spatial autocorrelation bias. In particular, the models without ρ e λ are normal fixed-effects panels; models with ρ ≠ 0 and λ = 0 are labeled as SAR (spatial autoregressive model); models with ρ ≠ 0 and λ ≠ 0 are labeled as SAC (spatial autoregressive combined model); models with ρ = 0 and λ ≠ 0 are labeled as SEM (spatial error model) (Lee and Yu 2010; Belotti et al. 2017).

Table 2 Summary statistics for stacked differences panel
Full size table

Table 2 shows the descriptive statistics of the variables used in the analysis. All the variables other than “exposure to robotization” come from the OECD Regional databaseFootnote 7. The trends of the five dependent variables considered in the four periods are reported in Appendix (Figs. 3, 4, 5, 6 and 7).

To proxy the role of policies, and being aware that we are simplifying the issue, we interact the regional exposure to robotization with a dummy taking the value of 1 for regions above the medianFootnote 8 in: (1) the level of investment in R&D, as a percentage of GDP; and (2) the level of education, as a percentage of the population between 25 and 64 years old with at least an upper secondary education levelFootnote 9.

The values of these dummies, and their respective classifications, are illustrated in Fig. 2. In this case, data for R&D expenditure and education level come from the OECD Regional database.

Fig. 2
figure 2

Regional dummies for high R&D expenditure and high education workforce. Source: authors’ original calculations on OECD Regional database and EUROSTAT database

Full size image

Figure 2 shows the existence of national patterns, which significantly affect regional characteristics and, consequently, related dummies. This is an expected result since national policies importantly influence R&D and education levels, even though a relevant difference can be noted. Specifically, as far as education is concerned, only national patterns arise: regions belonging to central and northern European countries show performances above the median, contrary to what happens in southern Europe. The same cannot be said unequivocally for R&D investment, since France, Spain, and Italy show some regional heterogeneity.

Finally, the fact that the two dummies differ in some, but relatively few, regions, suggests that similar results could be expected when using the two dummies in the regressions. For this very reason, however, the possible existence of differences in the estimation results can provide interesting insights about the relative role of the two policies.

5 Results

Tables 3, 4, 5, 6 and 7 report the outcome of the econometric analysis. Results are consistent regardless of the different variables and statistical technique used, confirming the different points of view reported in the theoretical section. Robotization negatively affects households’ income growth, whereas results on employment, unemployment, labor force participation rate, and labor utilization are less unequivocal. Yet, what is crucial here is the relation between EXP and long-term policies.

Table 3 Stacked differences panel, Dep. Var: (real) household income per–capita growth rate
Full size table
Table 4 Stacked differences panel, Dep. Var: Δ Employment rate
Full size table
Table 5 Stacked differences panel, Dep. Var: Δ labour force participation rate
Full size table
Table 6 Stacked differences panel, Dep. Var: Labor Utilisation (worked hours per capita) growth rate
Full size table
Table 7 Stacked differences panel, Dep. Var: Δ unemployment rate
Full size table

Table 3 reports the results for all models. Specifically, it shows that the household disposable income growth rate (the dependent variable) is negatively correlated with EXP in the case of regions with a low level of investment in R&D. The regions with high levels of investment in R&D show a better effect of EXP on the dependent variable with respect to the effect it exerts on the dependent variable of the region that invest less in R&D. However, this positive difference is not sufficient to cause an overall positive correlation between EXP and the household disposable income growth rate, and neither in the region with a high level of investment in R&D. In the empirical specification with controls, the overall coefficient of regions with a high level of expenditure in R&D (EXP+EXP * HIGH R&D) is never significantly different from zero.

Similarly, the results in the second part of the table, concerning the interaction between EXP and the educational level of the workforce (HIGH EDU), highlight how regions with lower levels of education suffer from a reduction in households’ disposable income growth rate, while those with higher levels of education experience better results. In the case of HIGH EDU, the overall correlation (sum of coefficients) is positive for regions with a higher share of educated labor force.

The education level appears to be more important than the R&D expenditure since the difference between regions with high and low level of education is more relevant (the coefficient of EXP * HIGH EDU is larger than the coefficient of EXP * HIGH R&D) and the overall coefficient (EXP+EXP * HIGH R&D) is significantly larger than zero in models that take into account controls and spatial correlation.

Table 4 shows the results for all models considering the employment rate as the dependent variable. In this case, EXP is not negatively correlated with the variation in the employment rate of regions with low levels of investment in R&D nor with a low level of educated workforce. The correlation between EXP and the variation in the employment rate is positive and significant in the regions with a higher propensity to invest in R&D and which have well-trained workforce, to the point that the overall correlation is always positive and significant (EXP+EXP * HIGH R&D and EXP+EXP * HIGH EDU significantly greater than zero). Again, the parameters related to differences in education level are larger than those related to R&D expenditure.

The results of Table 5, concerning the variation in the labor force participation rate, and Table 6, regarding the variation in the labor utilization, mutatis mutandis, are totally comparable to those concerning the employment rate (Table 4). Again, the parameters related to differences in education are larger than those related to R&D expenditure.

Finally, Table 7 suggests identical considerations about the correlation between EXP and variation in the unemployment rate due to variation in R&D expenditure and education level (with inverted signs). In general, the parameters introduced to control for possible spatial correlations (r and l) turn out to be significantly different from zero, confirming the appropriateness of the analysis when controlling for these aspects. Finally, the homogeneity of the results among the different model specifications supports the soundness of our methodology and results.

The correlation between EXP and the different dependent variables, focusing on the sole models with control, is summarized in Table 8Footnote 10. Four main stylized facts can be outlined. First, in the regions that invest more in R&D, we find a positive correlation between EXP and labor utilization, but not with the income growth rate, meaning that employees work more for the same overall wage. Secondly, in the regions with a higher level of educated citizens, we find a positive correlation between EXP both with labor utilization and income growth rate, meaning that employees work more for a higher overall wage. Thirdly, among these two patterns, that related to education and training appears to be more capable to cope with the effects of robotization, as proven by the higher value of the estimated coefficients and the positive association not only with the quantity of labor (as in the case of R&D), but also with income growth rate. Finally, all the variables considered are significantly associated with the difference between regions both with high and low level of education and R&D. Pushing such a result forward, we may argue that, in the lack of policy interventions, the distance between regions which invest more in education/R&D and those which invest less will increase as robotization expands.

Table 8 Correlation between regional EXP and variables under analysis (models with controls)
Full size table

Our results support the idea that more than a dichotomic positive or negative effect of robotization on employment and wages, we witness asymmetric outcomes depending on the relationship between the socioeconomic structure of a region and the policy implemented at its level. There may exist a complementarity effect between humans and robots as stated by many scholars. Nonetheless, the complementarity drawback may be improved by education and R&D policies, and by the subsequent positive social externalities they trigger. On the contrary, an increase of EXP in countries with low levels of R&D investment and education may amplify the negative consequences in terms of household income in the face of a stable trend of labor (employment rate, labor force participation rate, labor utilization) and unemployment rate. This means that the assumption of Pigou (1933 – point 4 of the positivist view), who argues that maintaining employment is possible by lowering wages, may be underway. However, this assumption, like all the others falling into the positivist view, basically focuses on the sole overall employment trend, regardless of the actual improvement in economic wellbeing at the individual and aggregate levels. Finally, results appear to confirm the polarizing nature of disruptive innovation, like robotization, which has been stressed by some authors, but which deserves more attention.

6 Conclusion

The current robotization process has renewed the long-lasting debate between techno-optimists and techno-pessimists about its effects on employment, and, ultimately, on citizens’ wellbeing.

Against this background, the main contribution of our work is to empirically show that this impact can be both positive and negative, and that this heterogeneity is correlated with the implementation and the effects of long run policies (such as, but not only, R&D and education). In so doing, we contribute to fill the gap between the theoretical foundations of such an approach and the related (but currently scarce) empirical evidence. Furthermore, by focusing on the regional dimension, we cope with the scarcity of analyses at this territorial level mainly due to the lack of data. Implementing the Acemoglu and Restrepo (2020) methodology to compute the Regional Exposition to Robotization in 150 NUTS-2 regions of the Euro area, we show that the related risk in terms of employment losses is negatively correlated with higher levels of regional investment in R&D and education. In other terms, regions that invest more in R&D and education are more able to cope with the risks related to robotization. This result supports the idea that high levels of R&D and education trigger positive socioeconomic externalities, which allow to exploit the complementarity effect with machines while mitigating the negative outcomes of the substitution effect. Since the levels of investment in R&D and of citizens’ education are importantly affected by public policies, one of the main issues to deal with regards the role of public intervention. Public policies must face the negative externalities arising in those regions with a suboptimal level of investment in R&D and education (which usually also suffer from poor economic performances) through direct intervention (public spending) and indirect intervention (incentives and subsidies to economic agents), especially in the absence of European solidarity mechanisms. Results appear to confirm the necessity of targeted R&D and education policies to avoid that robotization could further the ongoing polarization process triggered by technological advance.

 This issue has become crucial given the importance that regional and local levels of government have gained in recent decades, especially with regard to cohesion policies. Especially after the 2008 financial crisis, the divergence between European regions has widened, undermining the credibility of the European project. Such an increase in divergence can be attributed to both institutional factors, such as the austerity policies which have been implemented in Greece, Italy, and Spain (Perez and Matsaganis 2018), and economic ones, such as the structural change which has occurred in the last two decades. Although robotization has exerted a crucial role in boosting structural change, our results suggest it cannot be considered as the only cause of the increase in income and territorial polarization. Other factors should be tested in further research, some of which are strictly related to the institutional level, pertaining, for instance, to the reduced role of the states in the economy, the flexibilization of the labor market, and the decreasing capacity of policy makers to implement redistribution policies.

Finally, despite the focus on European regions, our results clearly underline the role of national policies. Education, in particular, which seems to be more able than R&D to turn into positive the effects of robotization, responds to national patterns. One of its main drawbacks, though, is that education policies usually produce positive effects in a longer time than R&D policies. This finally calls for the joint implementation of both of these policies.