Robots and reshoring: a comparative study of automation, trade, and employment in Europe

Abstract

This paper constitutes a comparative assessment of the impact of robots on local labor markets across eight European countries. Doing so, we find that robots generally reduce employment in the manufacturing sector, while their impacts on total employment are more ambiguous. Though local markets experienced significant employment losses in Italy, Norway, and the United Kingdom, we find no statistically significant impact of robots on employment in Germany, Sweden, Denmark, Finland, and Spain. Job losses were seemingly offset by employment gains in other sectors of the economy. We explore two mechanisms that might explain these patterns: differences in investment in complementary job-creating technologies and robot-induced reshoring. Our analysis provides some support for both mechanisms in elucidating the differential impacts of robots on jobs across space.

1. Introduction

How have workers in Europe fared from automation in recent decades? While some see robots as a harbinger of technological unemployment (Ford, 2015), the evidence is mixed so far. In the United States and the United Kingdom, there is compelling evidence that industrial robots have reduced employment (Acemoglu and Restrepo, 2020; Chen et al., 2022), but in Germany, jobs lost to robots in manufacturing were offset by jobs gained in other sectors (Dauth et al., 2021). Thus, even though robots reduce the demand for labor in the production of manufactured goods, they also boost productivity and might in some cases create new tasks (Acemoglu and Restrepo, 2018). Which effect dominates is likely to depend on specific labor market characteristics, complementary investments in job-creating technologies, as well as robot-induced reshoring.

In this paper, we examine the employment effects of industrial robots in eight European countries: Denmark, Finland, Germany, Italy, Norway, Spain, Sweden, and the United Kingdom. To our knowledge, this constitutes the first comparative assessment of the impacts of automation across local labor markets in Europe. For our analysis, we construct a measure of exposure to automation using data from the International Federation of Robotics (IFR), which provides annual robot counts across industries and countries. Thus, as in Acemoglu and Restrepo (2020), the variation in our automation measure stems from the fact that local markets within each country specialize in different industries, making some places more exposed to automation than others. Following Autor et al. (2013), showing that Chinese imports have had dramatic negative impacts on employment in local labor in the United States, we also report the effects of Chinese imports across European countries for comparison. For this analysis, we use the trade data from the UN Comtrade database.

A concern with our identification strategy is that robot adoption in a given industry might be related to other confounding factors affecting that particular industry. To mitigate this concern, we use the industry-level operational robot stock of other technologically advanced European countries as instruments for a country’s industry exposure.¹ Our ordinary least squares (OLS) estimates show that the impact of robots on employment is highly heterogeneous across Europe. Turning first to the manufacturing sector, we find that robots have a consistently negative impact on manufacturing jobs across all countries, though some coefficients are imprecisely estimated. This stands in contrast to imports from China, which have a positive impact on manufacturing employment in some Nordic countries, notably in Finland, while reducing employment in Denmark, Italy, Spain, and the United Kingdom.

The impact of robots beyond the manufacturing sector is more ambiguous. Though our analysis draws upon different data sources provided by national statistics offices for slightly different periods of time, and cross-country comparisons thus need to be made with caution, some patterns are nonetheless noteworthy. While local labor markets with greater exposure to robots experienced significant employment losses in Italy, Norway, and the United Kingdom, we find no statistically significant impact of robots on employment in Germany, Denmark, Finland, and Spain. Quantitatively, our baseline estimate suggests that the adoption of one more robot per thousand workers in a given local labor market reduced its employment-to-population ratio by 0.55 percentage points relative to other areas in Italy, by 1.93 percentage points in Norway, and by 0.39 percentage points in the United Kingdom. We note that the employment effects of robots in Norway are significantly larger in magnitude than previously reported estimates for the United States and other European countries (Acemoglu and Restrepo, 2020; Dottori, 2020; Dauth et al., 2021; Kariel, 2021).

Our findings add to a growing body of work, showing that the impact of robots on jobs has been highly heterogeneous across even advanced economies. While Graetz and Michaels (2018) find no evidence that robots reduced employment on average in a sample of 17 countries, there is compelling evidence that robots led to job losses in the United States (Acemoglu and Restrepo, 2020). In Germany, in contrast, job losses from robots in manufacturing were offset by an expansion of employment in services (Dauth et al., 2021), while robots seemingly boosted employment at the firm level in the Netherlands (Bessen et al., 2020). Though we similarly observe significant variation in our sample, we also note some general patterns. Across the investigated countries, robots seem to have reduced employment in the manufacturing sector, though we note that, in some cases, the effects are not statistically significant. We also find that the adverse employment consequences of robots are primarily borne by young and middle-aged workers in general and unskilled men in particular. Finally, we note that unlike other computer technologies, which tend to complement skilled labor (Autor et al., 2003; Autor and Dorn, 2013), the direct employment effect of robots seems to have been replacing unskilled labor without increasing the demand for skilled workers.

Second, we add to an emerging literature examining how the most recent wave of technologies, like robots, compares to previous waves (Lee and Lee, 2021), showing that new technologies, commonly associated with the fourth industrial revolution, are still deeply embedded within Information and communication technologies (ICT) from the previous generation (Martinelli et al., 2021). Building on the intuition that investments in robots might be bundled with investments in ICT in some places but not in others, we explore the heterogeneous impacts of robots on jobs across space along these lines. Doing so, we provide suggestive evidence that the positive impact of robots on jobs outside the manufacturing sector in Spain stems from higher levels of ICT investment. We also note that countries, where the displacement effect from robots is particularly large, exhibited the lowest levels of investment in job-creating technologies. In line with the framework of Acemoglu and Restrepo (2020), we argue that investment in enabling technologies, like ICT, which create new tasks to offset the displacement effect, is important to understanding employment patterns across countries.

Third, we build on a set of studies investigating the impact of automation on reshoring (Carbonero et al., 2018; Hallward-Driemeier and Nayyar, 2019; Faber, 2020; Kugler et al., 2020; De Backer and DeStefano, 2021; Bonfiglioli et al., 2022b). Notably, in the US context, Bonfiglioli et al. (2022b) find that the displacement effect of automation is less pronounced in cities that are more exposed to offshoring. To capture offshoring, we create a variable capturing the exposure to Chinese imports of intermediate goods, following Feenstra and Hanson (1999). Though most coefficients are imprecisely estimated, we note that robots induced reshoring in Germany. This helps explain the relatively muted impact of robots on jobs in the German context, also previously documented in Dauth et al. (2021).

The remainder of this paper is structured as follows. In section 2, we describe our data and provide some descriptive statistics. Section 3 outlines our empirical strategy, while section 4 discusses our results, including a host of robustness checks. Section 5 explores potential mechanisms underpinning the cross-country differences we observe, while section 6 outlines our conclusions and discusses avenues for future research.

2. Measurement and data

2.1 Robots

Following the empirical literature (Dauth et al., 2017; Graetz and Michaels, 2018; Acemoglu and Restrepo, 2020), we rely on robot data compiled by the IFR for our analysis. This allows us to measure the impact of advanced robotics technology on local labor markets from 1993 onwards. IFR (2014) defines industrial robots as “automatically controlled, re-programmable, and multipurpose” machines that are autonomous and that can be adapted to undertake a variety of tasks without the help of human operators. Their data have been collected from industrial robot suppliers around the world and consolidated at the industry-country level.

We proceed to construct a dataset with 12 disaggregated manufacturing industries based on IFR’s two-digit industrial classifications. These industries include foods and beverages; textiles (including apparel); woods and furniture; paper and printing; plastic and chemicals; minerals, glass, and ceramics; basic metals and metal products; industrial machinery; electronics; automotive; other transport equipment; and other manufacturing. We further consolidate the data into six additional broad industries for the following non-manufacturing sectors: agriculture, forestry and fishing; mining; utilities; construction; education, research and development; and services. A common concern in the literature is that there are still some robots that are not classified by the IFR. For example, more than 50% of robots in Denmark are unspecified before 2000 and around 5% of robots are unclassified in Spain. To account for this shortcoming, we allocate unspecified robots to each industry by its share of the remaining operational stock for each year following the approach of Acemoglu and Restrepo (2020).² Another caveat with using the IFR data is that our analysis does not take into account the impact of other automation technologies such as artificial intelligence or services robots on employment. Nonetheless, the IFR data, which provide consistent information on robot adoption across countries over time, have been employed widely in studying the employment impacts of automation technologies.

Figure 1 plots the evolution of the robot intensity in production (i.e., the operational stock of industrial robots per 1000 workers) for Denmark, Finland, Germany, Italy, Norway, Spain, Sweden, and the United Kingdom (panel A), as well as trends in the total operational robot stock by industry (panel B). We note that Germany has the highest robot intensity throughout the investigated period. Although some countries have been catching up, others have fallen behind. For example, in 1995, the German robot intensity is around 1.5 robots per thousand workers and 0.25 in Norway. By 2007, the German robot intensity had increased above 4, while it was still a mere 0.44 in Norway. To be sure, some of these differences can be explained by different patterns of specialization. As shown in panel B, the European automotive industry has adopted most robots over the past two decades, followed by chemicals, metals, foods, machinery, and electronics. The use of robots is much less common in other industries, especially in non-manufacturing sectors. To address the potential bias of our results being driven by events in the automotive industry, we examine the effect of automotive and non-automotive robot adoption on local labor markets in the robustness section. Due to the lack of robot usage data at the regional level, following the approach of Acemoglu and Restrepo (2020), we construct an adjusted penetration of robots (APR) to measure the industry-country level variation at different time periods, which is the change in robot installation per thousand workers with an adjustment for the industry-wide output expansion. We integrate the IFR data with the number of employees and real gross output at the country-industry level collected from the EU KLEMS: capital, labour, energy, materials and service (EUKLEMS) (November 2009 Release, updated March 2011; see van Ark and J¨ager, 2017). In our study, the measure of the APR at the industry-country level between t₀ and t₁ is defined in equation (1):

where $R_{i,t}^c$ is the number of robots in industry i in country c at time t, $g_{i,\left(t_0,t_1\right)}^c$ is the growth rate of real gross output (2007 = 100) of industry i in country c between $t_0$ and $t_1$⁠, and $L_{i,1990}^c$ is the baseline employee level (per thousand workers) in industry i in country c. $\frac{R_{i,t_0}^c}{L_{i,1990}^c}{g}_{i,\left(t_0,t_1\right)}^c$is the adjusted term which accounts for the fact that robot adoption could be confounded by an expansion of product demand in industry i.

Ideally, we want to construct the measure of robot penetration which only captures the exogenous technology shock. However, robot adoption may be well affected by local industry-specific demand shocks. To address this concern, we use the average APR in the same set of industries in seven other European countries. For Denmark, for example, the average APR is constructed by using the change in robots intensity and the growth of real gross output from Finland, France, Italy, Norway, Spain, Sweden, and the United Kingdom. This allows us to pick up the variation of robot adoption coming from other technologically advanced countries, which should not be correlated with shocks to local demand.³ Specifically, we construct our average APR variable as follows:

where $R_{i,t}^j$ is the number of robots in industry i in country j at time t, $g_{i,\left(t_0,t_1\right)}^j$ is the growth rate of real gross output (2007 = 100) of industry i in country j between $t_0$ and $t_1$⁠, and $L_{i,1990}^j$ is the baseline employee level (per thousand workers) in industry i in country j. J indicates European countries, including Denmark, Finland, France, Italy, Norway, Spain, Sweden, and the United Kingdom used to construct the instrumental variable, and the specific host country c is excluded in the sum of change in robot penetration. As we would expect, Figure 2 shows that our two APR measures, with or without the automotive industry, are highly correlated. In the automotive industry, for example, the APR indicates that the adjusted increase of industrial robots per thousand workers is between 1 and 4 across all eight European countries. Turning to the relationship between our APR measure and employment, Figure 3 highlights the negative relationship between our APR variable and employment growth across industries. We note that industries that installed more industrial robots typically saw a reduction in employment.

2.2 Chinese imports

Our trade data are collected from the UN Comtrade database with import and export data at the six-digit Harmonized System (HS) level across countries and years. Specifically, we use trade data for eight European countries as well as China from 1993 onwards.⁴ We do this to make our trade data line up with our data on robots. To have comparable industrial classifications across countries, we first map the six-digit import data onto three-digit SIC industries based on the HS1992-SIC crosswalk taken from Autor et al. (2013). Next, we aggregate the SIC industries into 12 manufacturing industries by the IFR industrial classifications, which are used to construct the robot exposure variable.

Figure 4 shows the import penetration ratio, which is defined as the share of the value of imports over total domestic demand, across the countries in our sample between 1995 and 2017. In the 1990s, the Chinese import penetration ratio is below one for all considered countries. However, after China joined World Trade Organization (WTO) in 2001, this ratio increased at staggering speed, reaching more than one in most countries in 2007, at the dawn of the Great Recession. We note that all countries in our sample have experienced rising import competition from China over the past two decades.

As in Autor et al. (2013), our APR is calculated as the change in Chinese imports per worker (per thousand US dollars, 2007 = 100) in a local labor market:

where $M_{i,\left(t_0,t_1\right)}^{China-c}$ is the change in imports from China (per thousand US dollars, 2007 = 100) in industry i between $t_0$ and $t_1$for country c, and $L_{d{t}_0}^c$ is total employees in local labor market d at t₀.

To address the concerns that the change in local employment and imports from China may be affected confoundedly by unobserved product demand shocks, we employ an instrumental variable strategy to mitigate the potential endogeneity from the supply-driven component of rising Chinese imports after China opened its market in the early 1990s and joined the WTO in 2001. Specifically, we instrument for the home country’s growth in Chinese imports by using the historical industry structure and the growth of imports from China in four other high-income (HI) countries, which also faced Chinese import competition during the same period. These countries include Australia, Japan, New Zealand, and Switzerland. This specification is written as follows:

where $M_{i,\left(t_0,t_1\right)}^{China-HI}$ is the change in the sum of Chinese imports of four other HI countries (per thousand US dollars, 2007 = 100) in industry i between $t_0$ and $t_1,$and $L_{d{t}_{-1}}^c$ is total employees in local labor market d prior to 1990. We use the historical employment data at the local labor market level to control for potential confounding labor demand shocks.⁵

Panel A of Figure 5 shows that the change in Chinese imports of the home country and the aggregate change of four HI countries are highly correlated. Electronics and textiles are the main industries exposed to import competition from China. Unlike our APR measure (Figure 3), we do not observe a negative correlation between import competition from China and employment growth across industries in panel B. This suggests that the impact of Chinese import competition on local employment might have been less pervasive in Europe relative to the United States (Autor et al., 2013).

2.3. Additional control variables

For our analysis, we collect data from national statistics offices and Integrated Public Use Microdata Series-International (IPUMS)-International to construct our long-difference specifications for the period between the early 1990s and 2007 at Nomenclature of Territorial Units for Statistics (NUTS) 3 region or a more granular level.⁶ To be comparable to the existing literature (Dauth et al., 2017; Acemoglu and Restrepo, 2020; Dottori, 2020), the main outcome variable is the employment-to-population ratio, including all employed persons/employees across all sectors. Table 1 provides data coverage of each country, and  Appendix A documents the data collection process in more detail.

As additional covariates, we include a set of demographic variables, including the log of population, the male population share, the share of the population above the age of 65 years, the share of the population that is foreign-born, as well as the share of the population with a higher education degree.⁷ We also include the ethnic population share when the data are available. Furthermore, by including a set of local industry characteristics, we control for the confounding effects of the pre-existing industry structure on local employment. These industry variables are the share of employment in light manufacturing, the share of employment in mining, the share of employment in construction, and the share of female workers in manufacturing.⁸ The tables with complete descriptive statistics for each country can be found in  Appendix A.⁹

3. Empirical strategy

3.1. Exposure to robots and Chinese imports

Since we cannot observe the actual robot usage in a local labor market, following Acemoglu and Restrepo (2020), we use a shift-share design to apportion each industry’s robot penetration across local labor markets based on their industrial employment shares. To capture the characteristics of local labor markets, we use units that roughly correspond to NUTS 3 or a more granular level. For most countries, our study period ends in 2007 to alleviate potential unobserved shocks confounded by the Great Recession and Brexit.¹⁰ Local industry employment data are collected at the International Standard Industrial Classification (ISIC) two-digit level from the relevant national statistics offices or the Integrated Public Use Microdata Series International (IPUMS-International, Minnesota Population Center, 2020).¹¹

To analyze the impact of industrial robots on employment, our empirical identification strategies rely on substantial variation in local industry specialization. This means that local economies that have specialized in industries where more industrial robots are installed should be differentially affected by the robot revolution within a country. The exposure to robots variable, thus, measures the predicted instead of actual change in the number of robots in each local labor market. A remaining concern of our identification strategy, however, is that shocks to local labor demand, such as a local recession or changing tax incentives, are affecting the adoption of robot technology. To address this, we take the employment shares from the previous decade, which capture the historical distribution of industrial employment before industrial robots were adopted in the local market. Equation (5) defines the instrument for the exposure to robots, accounting for unobserved industrial and local demand shocks, as follows:

where ${\overline{APR}}_{i,\left(t_0,t_1\right)}$is derived from equation (2), and $l_{di}^{t_{-1}}$ allows us to control for the anticipation of the adoption of robotics technology in the early 1990s.¹²

Similar to the measurement of robot exposure, our exposure to Chinese imports is the change in the sum of Chinese imports of four other HI countries multiplied by its share of national industry employment prior to 1990:

where ${\overline{API}}_{id,\left(t_0,t_1\right)}$ is defined in equation (4) and $\frac{L_{id{t}_{-1}}^c}{L_{i{t}_{-1}}^c}$ is local labor market d’s share of national employees in industry i at $t_{-1}$⁠. Again, we use the historical industrial employment composition at the local labor market level to alleviate the concern of anticipation effects on the growth of Chinese imports.

Our baseline specification is a long-difference OLS to estimate the impacts of robots and trade on local labor markets during the period t₀ to t₁:

where the outcome variable $Y_{d,\left(t_0,t_1\right)}^c$ is the change in the employment-to-population ratio between $t_0$ and $t_1$ in local labor market d in country c. $E{R}_{d,\left(t_0,t_1\right)}^c,$our main variable of interest, is the exposure to robots, while $EC{I}_{d,\left(t_0,t_1\right)}^c$ captures the exposure to Chinese imports. These variables are defined in equation (5) and (6), respectively. We also include regional fixed effects, ${\delta }_r$⁠, controlling for time-invariant trends across regions, and a vector of control variables, ${\mathbf{X}}_{d{t}_0}$⁠, measuring the start-of-period demographic and industry characteristics. The covariates included for each country are documented in  Appendix A.

3.2. Validity checks

An underlying assumption of our OLS specifications is that local labor markets specialized in industries which have adopted automation technology more rapidly have been more exposed to industrial robots than others and have not experienced other local demand shocks. In other words, unobserved local demand shocks might occur as threats to this assumption. In this section, we scrutinize potential validity threats to our identification strategy.

Table 2 provides summary statistics indicating how local labor market characteristics differ across four quartiles of exposure to robots. (For the full set of variables, see  Appendix A.) Column 1 shows the mean for all local labor markets, while columns 2 to 5 present the means of main outcome and exposure to Chinese imports variables by quartiles of exposure to robots. We note that in columns 2 to 5, the trends between exposure to robots and Chinese imports are not consistent across the four quartiles among the eight European countries, while the labor markets that were more exposed to robots experienced more negative labor market trends.

We next explore the variation across manufacturing industries, where the robots and Chinese imports are highly concentrated, for adjusted robot exposure, change in imports from China, as well as the growth of ICT, computer software, research and development, and fixed capital for the European countries. Figures A1–A8 in  Appendix A plot the normalized scale with the largest number for each indicator across 12 manufacturing industries. We note that industries with the highest robot penetration, such as automotive or chemicals and plastics, are different from those highly exposed to imports from China like electronics or textiles. This indicates that we should be able to distinguish the effects of exposure to robots from Chinese imports in our main specification. Moreover, those industries adopting more robots seem not to be correlated with other indicators such as ICT or fixed capital investment in the industry level. This supports our identification strategy that our exposure variables are not confounded with other factors that drive the demand shock in the local labor markets.

Furthermore, as our main measurements of both exposure of robots and Chinese imports are shift-share variables, there exits concern that the variation of the regressors is just driven by a handful of industries. Indeed, Figures 2 and 5 show that the robot adoption and imports from China are concentrated in a few industries, such as automotive or electronics. Hence, based on Goldsmith-Pinkham et al. (2020), we construct Rotemberg weights to identify the relative contribution of each industry to the overall explanatory power of the shift-share instrument. Doing so, we find that a small number of industries account for a large share of the Rotemberg weights across European countries. For example, in Sweden, automotive alongside chemicals and plastics have the highest Rotemberg weights and explain 90% of the predicted power of the Bartik shift-share robot exposure variable, while electronics and chemicals and plastics account for 86% of exposure to Chinese imports.

According to Goldsmith-Pinkham et al. (2020), the identification of Bartik style instrument is valid if the underlying local employment share is not correlated with other local factors which have impacts on the outcome variable. Thus, we exploit the relationship between the top two local industrial employment shares and start-of-period covariates in the local labor markets that could affect our outcome variables. Besides, we also regress our variables of interest, exposure to robots and Chinese imports, on those local covariates. Tables 17–24 in  Appendix A present the OLS estimates for industries with the largest Rotemberg weights as well as the exposure to robots and Chinese imports.¹³ We find that some covariates have significant effects on local industrial employment, and hence, they are included in our main specifications to account for potential common trend shocks.

Finally, a remaining concern is that some industries could have suffered declines in employment before the implementation of industrial robots or that local labor markets which adopted more robots might have experienced other local negative shocks. To that end, we include the lagged change in the employment-to-population ratio on the right-hand side in equation (7) to directly control for potential employment pre-trends.

4. Results

Table 3 presents our results from estimating equation (7), where we regress the change in the total employment-to-population ratio on the exposure to robots in panel A and both the exposure to robots and Chinese imports in panel B.¹⁴ The specifications are weighted by population in the start-of-period to account for the variation in market size, and robust standard errors are used to control for the heteroskedasticity across regions. We drop singleton groups in regressions where fixed effects are nested within clusters. This helps to prevent overstating statistical significance and causing incorrect inference.

In panel A, we explore the relationship between the exposure to robots and the total employment-to-population ratio across countries. For Italy, Norway, and the United Kingdom, we find that the exposure to robots has a significant negative impact on employment in local labor markets with a coefficient of −0.52 (standard error = 0.14), −2.1 (standard error = 0.63), and −0.47 (standard error = 0.19), respectively. In the remaining countries, however, the relationship is imprecisely estimated.

In panel B, we further account for the impact of Chinese imports on local employment. Our estimates of the impact of robots remain significant and negative in Italy, Norway, and the United Kingdom after controlling for trade exposure to China. Quantitatively, the estimate of exposure to robots in Italy in panel B, for example, indicates that the adoption of one additional robot per thousand workers in a local market led to 0.55 percentage points lower employment-to-population ratio relative to other areas.

To explore these relationships further, Figure 6 provides a residual regression plot showing the variation of the exposure to robots. The solid line shows the regression estimates conditional on covariates included in panel B of Table 3, while the dashed line presents the same regression relationship excluding the top one percent of local labor markets with the highest exposure to robots or Chinese imports. The size of each circle indicates the local market’s population in the start-of-period, and we can observe a substantial variation in industrial composition across local labor markets within each country. The distribution of exposure to robots is mainly skewed to the right with only a handful of local labor markets with large values. Notably, many of them are specialized in automotive, such as Turin in Italy, Valladold in Spain, Olofström in Sweden, and Solihull in the United Kingdom, or the plastic and chemical industry in the case of Porvoo in Finland and Perstorp in Sweden. These highly specialized markets also reflect the industry-level variation in robot penetration in Figure 2. Hence, Figure 6 depicts countries with relatively many local labor markets exposed to robots, like Italy or the United Kingdom, which faced a larger displacement effect than others, such as Denmark or Spain.

4.1. Composition effects

Table 4 presents the estimated impact of robots and Chinese imports on the employment-to-population ratio in manufacturing in panel A as well as outside manufacturing in panel B. In the manufacturing sector, we find that robots reduced employment in all countries in our sample, but the effect is only statistically significant in Italy, Spain, and the United Kingdom. The impact of Chinese imports, in contrast, is not consistently negative but also not statistically significant in most countries. For example, we find that Chinese imports reduced employment in Italy, Spain, and the United Kingdom but seemingly boosted manufacturing employment in Finland.

Outside the manufacturing sector, we also find that robots increased employment in Spain. This is consistent with a reallocation of employment from manufacturing to non-manufacturing industries due to employment spillovers. For instance, robotization in manufacturing industries might increase productivity and the demand for complementary services such as engineering, consulting, and marketing. This countervailing effect might explain why both exposure measures do not have an impact on the aggregate employment-to-population ratio presented in column 7 of Table 3. Conversely, robot exposure reduced the local demand for jobs outside the manufacturing sector in Germany, Italy, and Norway. We note that our estimates differ from those of Dauth et al. (2021), who find that robotization increased employment outside the German manufacturing sector between 1994 and 2014. One possible explanation for this difference is that robots did more to offset the displacement effect in the post-recession period, which we do not consider due to the potential confounding effects of the recession itself. For example, Jungmittage and Pesole (2019) show that robots in Europe had a much smaller impact on aggregate labor productivity in the 1995–2007 period, relative to the period 2008–2015, when robots spread to more industries. This might also explain why Dottori (2020) finds no harmful impact of robots on total employment across Italy between the early 1990s and 2016. Reassuringly, however, for the period up until 2001, his estimates are similar to ours.

Finally, it is possible that automation and import competition are substitutes in some countries, meaning that the adoption of robotic technology could be associated with a reduction in imports, thereby counteracting job losses domestically. To that end, we include an interaction between exposure to robots and Chinese imports in the baseline equation (7). Panel C of Table 4 presents the estimated interaction effect on manufacturing employment. We observe that the coefficient of the interaction variable is positive and estimated precisely in Finland, Germany, and Norway, suggesting that in local markets with higher exposure to Chinese imports, robots mitigated job losses in the manufacturing sector. This, we note, might partially explain the positive effect of exposure to Chinese imports on manufacturing employment in Finland documented in panel A. The extent to which robots induce import substitution, however, is highly heterogeneous across countries. In Denmark, Italy, and Spain, for example, the interaction is negative, though statistically insignificant, suggesting that the introduction of robots might induce more imports of related components. Building on these intuitions, we further explore the role of robot-induced reshoring in Section 5.2.

4.2. Effects by automotive and other industries

As shown in Figure 1, the automotive industry has experienced substantial growth in robot adoption relative to other industries since 1993. This skewed robot usage could cause biased estimates driven by this particular industry. In Table 5, we address this concern by differentiating between the exposure to robots in automotive, all other industries, and highly exposed manufacturing industries. Thus, our automotive exposure variable measures the robot adoption in the automotive industry, while the other variables capture the use of robots in all other industries. The estimates in panel B show that the effects of robot adoption in other industries are generally negative and significant across most countries, while in panel A, the effects of exposure to robots in the automotive industry are imprecisely estimated. We also examine the effect of exposure to robots in highly exposed manufacturing industries, which include automotive, plastics and chemicals, metal products, basic metals, electronics, and food and beverages. Panel C shows that the effects are generally negative and significant in Italy, Norway, and the United Kingdom. These results suggest that the displacement effects in local labor markets are not mainly driven by the automotive industry but by other industrial robots across European countries. As shown by Jungmittage and Pesole (2019), European industries which deploy more robots are also more productive. Possibly, the automotive industry, which was the very first industry to adopt robots, might have benefited from relatively large productivity gains that canceled out the displacement effects within the sector. Based on the French firm-level data, Bonfiglioli et al. (2022a) find that the expanding firms may adopt more robots that could also offset the jobs lost across industries.

4.3. Impacts by demographic groups

We next turn to investigating to what extent different demographic groups in the labor market has been impacted differentially by robots. Table 6 reports OLS long-difference specifications of the change in the employment-to-population in response to robot exposure. Overall, in panel A, we find that the impact on male employment is consistently negative across the countries in our sample, though it is only statistically significant in Italy and the United Kingdom. This speaks to the fact that roughly 60% to 70% of the manufacturing workforce across the eight European countries in our sample is male. The picture for female employment is more mixed, though also imprecisely estimated, with the exception of Germany where the effect is negative.

In panel B, we explore the impact of robots across different age groups. On balance, robots seem to reduce the employment prospects of primarily younger workers. While imprecisely estimated, robots have a consistently negative impact on employment for those aged 24 years and younger, Italy being the exception. We also note that Acemoglu et al. (2021) suggest that middle-aged workers have been most exposed to automation in the United States, and in our sample the impact on middle-aged workers (aged 25–54 years) is more mixed: while middle-income workers in Italy and Sweden, where the automotive industry is most prominent, experienced employment declines, we find the opposite in Denmark, which is more specialized in light manufacturing, and where automotive industries are relatively small, as shown in Figure 2.¹⁵ Conversely, we find that robots increased employment among those aged 55 years and above in some countries, notably in Finland and Germany. One possible explanation is that automation increases the demand for supervisors and managers in some settings, especially when investment in robots is bundled with investment in ICT—an issue to which we shall return in Section 5.

Finally, in panel C, we explore the effect of robots on different skill groups, where we follow other studies (Autor et al., 2015; Acemoglu and Restrepo, 2020; Dauth et al., 2021), using educational attainment as a proxy for worker’s skill level. We define skilled workers as those with a college/university degree or above.¹⁶ Doing so, we find no evidence suggesting that industrial robots directly complement skilled workers, unlike other computer technologies (Autor et al., 2003; Autor and Dorn, 2013). An exception to this pattern is Spain, where robots increased the demand for skilled workers—and as we shall see, in Spain, investment in robots was seemingly more bundled with investment in ICT. However, this increase seems to have taken place outside of the manufacturing: as shown in Table 4, robots reduced manufacturing employment in Spain but increased employment in other sectors. Unsurprisingly, we find that unskilled workers are more likely to have seen vanishing employment opportunities due to robots, notably in Germany and the United Kingdom. We note that our findings are broadly consistent with the cross-country-industry results from Blanas et al. (2020), implying that young and unskilled workers have been most adversely impacted by robots.

5. Mechanisms

As outlined in Acemoglu and Restrepo (2020), the displacement effect of robots might be counteracted by either higher rates of productivity growth or investments in enabling technologies, which create new tasks for labor. While we are unable to fully disentangle heterogeneous impact of robots on employment we observe across countries, in the below, we seek to shed some light on potential mechanisms in two ways. First, building on Acemoglu and Restrepo (2020), who document a positive relationship between ICT capital and employment in the US context, we explore the potential countervailing role of ICT in creating new work. Second, we go beyond the existing literature in considering also the possibility for reshoring. Indeed, if companies can use robots rather than labor, they might seek to reshore production and save transportation costs.

5.1. The role of ICT

A growing body of work examines the distinctiveness of technologies associated with the so-called fourth industrial revolution (Martinelli et al., 2021) exploring their departure from the ICT technologies of the third industrial revolution (Lee and Lee, 2021). Based on patent data, these studies show that the more recent wave of technologies, like robots, are still deeply embedded within technologies from the previous generation.

However, the extent to which they are being embedded in production might differ across factories and locations, which in turn might affect our estimate outlined earlier. To explore this, we collect data on ICT equipment and computer software from the EUKLEMS 2019 release (Stehrer et al., 2019) as well as from Norway’s statistics office. To be compatible with our robot exposure measurement, we construct Bartik shift-share variables which are the long difference of the aggregate stock between 1995 and 2007, multiplied by the base-year employment share across local labor markets. Table 7 shows that the coefficients of exposure to robots are generally unchanged with similar magnitude and significance when pitched against those in panel B of Table 3. This suggests that robots have independent impacts on employment.

Yet, though the impact of robots is still mostly distinct, it becomes positive in Spain, suggesting that there might be differences in the bundling of robots and ICT across space. Notably, both ICT and software have significantly positive effects on local employment in Spain, which might also have served to offset the displacement effect of robots and contributed to job increases in the non-manufacturing sector as shown in Table 4.¹⁷ The negative effect of ICT on employment in Finland is probably due to the demise of Nokia in this period. We also note that in Figure 7 investments in ICT have been particularly low in Italy and the United Kingdom, where the adverse impacts of robots on jobs have been felt keenly. These findings speak to the importance of enabling technologies in offsetting the displacement effect, as highlighted by Acemoglu and Restrepo (2020).

5.2. Robot-induced reshoring

A further possibility is that robot adoption might have induced reshoring in some settings. Indeed, as shown by Bonfiglioli et al. (2022b), the displacement effect of automation is smaller in the local markets that are more exposed to offshoring in the United States. To that end, following Feenstra and Hanson (1999), we construct a variable capturing the exposure to Chinese imports of intermediate goods as a proxy for offshorability. The country-industry level data on imports of intermediate goods are collected from OECD Trade in Value Added (TiVA) for the period between 1995 and 2007. Similar to our measure for the exposure to Chinese imports in equation (3), the exposure to intermediate goods is calculated as the change in the stock of imports multiplied by the local industrial employment share at the start-of-period.

To examine whether robot adoption has triggered reshoring in some countries, we include an interaction term between robots and Chinese imports of intermediate goods in the baseline equation (7). Table 8 presents our results from estimating the impact of exposure to intermediate goods on manufacturing employment. The estimate of the interaction term is significant and positive for Germany, indicating that robots induced some reshoring in local markets that were more exposed to offshoring from China, plausibly offsetting some ongoing offshoring in the robot-reliant auto industry. We note that Finland, Norway, and the United Kingdom also show positive robot-induced reshoring effects, though they are not estimated precisely. The extent to which robots induce reshoring, in other words, varies significantly across countries. One plausible reason is that Chinese imports might substitute for domestic production in some instances and complement it in others. Indeed, we note that the interaction is again negative, though insignificant, for Denmark, Italy, and Spain—as with import substitution from China, as documented in Section 4.1. Overall, the sum of robot exposure coefficients shows that while the adoption of robots still reduces manufacturing jobs across European countries, the overall effects are much smaller or even muted in Germany and the United Kingdom.

6. Conclusions

In this paper, we examine how workers have fared from industrial robots and import competition from China in eight European countries: Denmark, Finland, Germany, Italy, Norway, Spain, Sweden, and the United Kingdom. Overall, we find that robots have reduced employment in the manufacturing sector, while the impacts on local labor markets, which also take into account indirect employment effects, are more ambiguous. Local markets with greater exposure to robots experienced significant employment losses in Italy, Norway, and the United Kingdom. The coefficients for the remaining countries are imprecisely estimated, partly because employment losses in the manufacturing sector were offset by gains outside the manufacturing sector, most notably in Spain. Unlike other computer technologies, which complement skilled workers in production (Autor et al., 2003; Autor and Dorn, 2013), we find that robots, which do not require an operator, had no significant impact on the demand for skilled workers. Spain is an exception, but also here, the increase in skilled employment seems to have taken place outside the manufacturing sector. We show that these patterns are related to bundled investments in enabling ICT technologies which helped offset some of the adverse employment consequences of robots, again notably in Spain.

While we are unable to fully disentangle the factors underpinning the differential impacts of robots on employment across countries, our findings seem to be driven by factors beyond variation in labor market institutions. For example, though there is evidence that robots have reduced employment in the United States (Acemoglu and Restrepo, 2020), employment losses in the German manufacturing sector were offset by job creation in other sectors (Dauth et al., 2021). Dauth et al. (2021) suggest that the relative strength of German trade unions might explain the differential impacts of robots on jobs in the United States and Germany.¹⁸ However, our findings show that countries like Norway, with a relatively high union density, experienced significant employment declines as robots proliferated across the country. Indeed, even though the Nordic countries have similar labor market institutions, they have fared differentially from automation. Rather, countervailing investments in job-creating enabling technologies and robot-induced reshoring seem to better explain some of the cross-country differences we observe, albeit to varying degrees. We deem disentangling the diverse impacts of robots on jobs in general, and its interaction with trade in particular, a fruitful avenue for future research.

Funding

This work was supported by the European Union’s Horizon 2020 research and innovation program under grant agreement no. 822330. C.B.F. also gratefully acknowledges funding from Citi and the Dieter Schwarz Foundation.

Footnotes

References

A.1 Denmark

Employment data for the years 1994 to 2000 were acquired from Denmark Statistics. Our analysis uses total employees at the place of work aggregated to ISIC two-digit industries and 99 municipalities (2007 version). Other local demographic characteristics and employment data are also collected from Denmark Statistics for the years 1984, 1994, 2000, and 2007 and were mapped using the municipality codes of the 2007 version for the years 1984, 1994, and 2000. Table 9 presents some summary statistics of outcome variables, controls and covariates.

A.2 Finland

Employment data for the years 1987 to 2007 were acquired from Finland Statistics. Our analysis uses total employees at the place of work aggregated to ISIC two-digit industries and 70 sub-regions. Other local demographic characteristics and employment data are also collected from Finland Statistics, where they provide data with a consistent geographical unit for the years 1987, 1993, 2000, and 2007. Table 10 presents the summary statistics of outcome variables, other variables of interest, and covariates.

A.3 Germany

We use data from the anonymous Establishment History Panel, 1975–2018. Data access was provided via on-site use at the Research Data Centre Forschungsdatenzentrum of the German Federal Employment Agency Bundesagentur für Arbeit at the Institute for Employment Research Institut für Arbeitsmarkt- und Berufsforschung, (IAB) and/or remote data access.¹⁹ The dataset covers all employees in the German labor market subject to social security, going back to 1975 for West Germany and 1992 for East Germany. The data encompass detailed information on the composition of employment and average daily wages, including consistent industry codes and demographic characteristics such as age, gender, and qualification. Our analysis uses total employees aggregated to ISIC two-digit industries and 402 districts (Landkreise and kreisfreie Staedte) for the years 1985, 1995, and 2007. We also construct data of employment by demographic groups using Betriebs-Historik-Panel (BHP) (1995–2007). Population data are collected from the German Federal Statistical Office for the years 1995 and 2007, where 1995 is the earliest available year at the district level. Due to the data availability from the German Federal Statistical Office, we have only 359 districts with population data. We also only report variables with at least 20 establishment observations from the Establishment History Panel. Table 11 presents the summary statistics of outcome variables, other variables of interest, and covariates.

A.4 Italy

Employment and industry data are collected from the Firm Census (Censimento generale dell’industria e dei servizi) for the years 1981, 1991, 2001, and 2011. Our analysis uses total employees from local business units, which include business, public, and non-profit institutions, aggregated to ISIC two-digit industries and 110 provinces (2009 version). Other local data are collected from the Population Census for the years 1981, 1991, 2001, and 2011 and aggregated at the province level by Istat in the Statistical Atlas of municipalities (Atlante statistico dei comuni). Demographic data are derived from the Population Census (1991–2011). Table 12 presents the summary statistics of outcome variables, controls, and covariates.

A.5 Norway

Employment and industry data for the years 1995, 2000, and 2007 are collected from Norway Statistics. Our analysis uses total employed persons at the place of work aggregated to ISIC two-digit industries and 74 economic regions (2018 version). Other local demographic characteristics and employment data are also collected from Norway Statistics, from which we aggregated detailed municipality codes into economic regions (2018 version) for the years 1995, 2000, and 2007. Employment data by demographic groups are only available after 2000. Table 13 shows the summary statistics of outcome variables, variables of interests, and covariates.

A.6 Spain

Local industry and demographic data are from the Population Census provided by IPUMS-International (Minnesota Population Center, 2020), for the years 1981, 1991, 2001, and 2011. IPUMS-International has provided data with consistent geographical boundaries, industry codes, and education classifications across census years. Our analysis uses total employment aggregated to ISIC two-digit industries and 50 provinces (NUTS 3 level) excluding overseas regions (Table 14).

A.7 Sweden

Employment and industry data based on administrative sources Registerbaserad arbetsmarknadsstatistik, for the years 1985 and 1990 to 2007, are from Sweden Statistics. Our analysis uses the total gainfully employed population at the place of work aggregated to ISIC two-digit industries and 100 local labor markets (1998 version). Other local demographic characteristics and employment data are also collected from Sweden Statistics, where the data are provided with consistent geographical units for the years 1985, 1993, 2000, and 2007. Table 15 presents some summary statistics.

A.8 United Kingdom

Employment and industry data are collected from the Business Register and Employment Survey (BRES) provided by National Online Manpower Information System (NOMIS), which entail aggregated local labor market data for consistent geographical units. Our analysis uses total employees aggregated to ISIC two-digit industries in 352 local authority districts (prior to the April 2015 version) covering England, Scotland, and Wales. Other local data are collected from Population Census for the years 1981, 1991, and 2001, provided by NOMIS. Employment by demographic groups is constructed by using BRES for gender (1991–2007), while age and skill groups (1991–2011) are taken from Population Census. For the wage income analysis, weekly pay by gender is taken from the Annual Survey of Hours and Earnings, provided by NOMIS, for the years 1991 and 2007. Table 16 presents some summary statistics of relevant variables.

Appendix figures