The labour market impact of robotisation in Europe

This paper explores the impact of robot adoption on European regional labour markets between 1995 and 2015. Specifically, we look at the effect of the usage of industrial robots on jobs and employment structures across European regions. Our estimates suggest that the effect of robots on employment tends to be mostly small and negative during the period 1995–2005 and positive during the period 2005–2015 for the majority of model specifications. Regarding the effects on employment structures, we find some evidence of a mildly polarising effect in the first period, but this finding depends to some extent on the model specifications. In sum, this paper shows that the impact of robots on European labour markets in the last couple of decades has been ambiguous and is not robust. The strength and even the sign of this effect are sensitive to the specifications, as well as to the countries and periods analysed.


Introduction
Few topics have historically attracted so much interest from the general public and the research community alike as the impact of technology on work (Mokyr et al., 2015). Jointly with globalisation, automation is one of the main sources of economic concern for European citizens, with almost three out of four considering that this phenomenon implies net job losses on the Continent (European Commission, 2017). Nevertheless, so far, this kind of technological anxiety may have led to an overestimation of the negative effects of automation. Productive technologies have in fact mostly contributed to raising the living standards of modern societies (Atack et al., 2019;Autor and Salomons, 2018). Nevertheless, the current debates often focus on whetherand to what extent -'this time is different' and whether current technological innovations are in fact more disruptive than those in the past (Autor and Salomons, 2018;Mokyr et al., 2015).
The aim of this paper is to explore the impact of a very specific technology, industrial robots, on jobs and employment structures in Europe during the period 1995-2015. We examine region-level variation in the exposure to robotisation, combining information on the deployment of this sort of capital by countries and sectors from the International Federation of Robotics (IFR) with region-level labour market statistics obtained from different Eurostat datasets.
We find that the impact of robot adoption on local labour markets depends on the time period analysed: whereas there is evidence of negative effects of the increase of robot density on employment in the period 1995-2005, the impact in the period 2005-2015 is positive or null. However, we perform extensive robustness checks and find that these results are often sensitive to the exact specifications of the estimations.
Regarding the effect of robot adoption on occupational change, our results are less clear. In some of our specifications, we identify a negative effect on the middle of the distribution or a positive effect on the extremes in the first period (which would be consistent with the hypothesis of a polarisation effect), but the differences between the change in the middle and the extremes are relatively small and often not statistically significant. In fact, for certain model specifications we find a more pronounced positive effect of robotisation on the middle tercile over the entire period, so this finding has to be interpreted with caution.
The impact of automation on the labour market is complex (Acemoglu and Autor, 2011;Restrepo, 2018, 2019). Firstly, the new technology might displace labour from those tasks it is brought in to perform. Secondly, it might involve the creation of new tasks and, consequently, foster the creation of employment associated with them. The same applies to those (pre-existing) tasks that are complementary to the technology introduced, the so-called productivity effect. In principle, it is unclear which effect dominates. Finally, Acemoglu and Restrepo (2018) add that, in the long run, the productivity effect becomes larger, since automation raises the rental rate of capital, which triggers further capital accumulation until a point at which the price of this factor reaches its steady state level.
There is an extensive body of literature aiming to disentangle the labour market consequences of technological changes, mainly focused on the task-content of jobs and its effect on different segments of the labour force (for the general developments regarding this topic, see, among many others, Acemoglu and Autor (2011) and Barbieri et al. (2020)). More specifically, there are several recent papers exploring the labour market implications of robot adoption on both developed and developing economies, with somewhat diverse conclusions. Exploiting sector-and country-level variation, Graetz and Michaels (2018) document a null effect of robotisation on jobs (with a negative effect on low-skill employment) and a positive effect on wages for the period 1993-2007 in developed countries. Klenert et al. (2021) show that the impact on employment is often positive when looking solely at manufacturing and extending the period of analysis to 2015. Carbonero et al. (2018), using a similar approach but looking at a larger sample of countries during the period 2000-2014, report a negative impact on employment, particularly in developing countries. Looking at a similar time window, De Backer et al. (2018) find a positive impact of robots on jobs in developed countries in some periods and a null effect for developing countries. For the US, Borjas and Freeman (2019) find a negative impact on wages and employment in the period 1996-2016, whilst Dahlin (2019) reports a positive impact on high-and middle-skill jobs between 2010 and 2015.
Another stream of literature exploits regional variation in robot exposure, assigning the increase in the stock of this technology by sector to subnational territories on the basis of the distribution of employment at the beginning of the observation period. This approach is adopted by , Dauth et al. (2021) and Chiacchio et al. (2018) when exploring the implications of robot adoption on employment in the US, Germany and a set of 6 European Union countries, respectively. Acemoglu and Restrepo (2020) report a negative impact of robotisation on both employment and wages. Chiacchio et al. (2018) find similar effects for Europe. Finally, in the work of Dauth et al. (2021), robot adoption seems to exert a negative effect on manufacturing employment and wages, but the overall effect on the labour market is null. This suggests that nonmanufacturing sectors are absorbing the employment made redundant in industrial activities and profiting from the productivity rise.
Finally, several recent studies based on firm-level data come to different conclusions. Whilst Moll and Lerch (2016), making use of the European Manufacturing Survey comprising firms in six European countries, find that the use of robots and employment are uncorrelated, Domini et al. (2021) and Koch et al. (2021), who make a greater effort than the latter study in identifying causality, find a positive effect of automation technology on employment creation. One exception is , who find that for the case of France, robots are associated with a decrease in employment on aggregate, as the reduction in labour outweighs employment creation effects driven by increased productivity.
The contribution of this paper is twofold. Firstly, we assess the robustness of previous studies on the impact of robotisation on local labour markets, including in our analyses a wider set of countries and time periods. Secondly, we explore the differential impact of robot adoption on different parts of the employment distribution (beyond educational levels), which allows us to examine potential links between robotisation and labour market polarisation.
The paper is structured as follows. In the second section, we describe in detail the databases employed in the analysis and the methodology used for exploring the impact of robot adoption on European labour markets. The third section presents the results of our econometric analysis. Finally, Discussion summarises and discusses the main conclusions of the paper.

Data
In order to assess the effect of robotisation on European labour markets, we combine several databases for the years 1995, 2005 and 2015. Regarding robot adoption, as many recent studies, we draw on the data from the World Robotics Database, specifically from the 2017 edition (IFR, 2018). This dataset, administered by the main association of producers of robots worldwide, contains information on industrial robot stocks and deliveries all over the globe. The concept or robot included in the IFR database is narrow. It comprises industrial machinery, digitally controlled, mainly aimed at handling operations and machine tending, welding and soldering and assembling and disassembling. In terms of accounting, robots are not part of information and communication technology capital (ICT), with the exception of the software needed to manage them. IFR robot data are plagued with many relevant problems (Bekhtiar et al., 2021;Fernández Macías et al., 2021). In order to deal with them, we employ the procedures of cleaning and imputation described by Fernández-Macías et al. (2021).
The labour market information used in this paper comes from several different sources. The first and main source is the European Union Labour Force Survey (EU-LFS), carried out jointly by the EU member states and Eurostat. The EU-LFS micro-data are a key source of information on employment figures and the main demographic characteristics of population across regions. Unfortunately, because of confidentiality and anonymity concerns, this dataset does not offer a detailed disaggregation by economic activityinformation effectively collected in the questionnaires. We have obtained the detailed distribution of the labour force by region and industry through several ad hoc requests to the Eurostat User Support. 1 Nevertheless, we resort to other databases in order to fill some gaps in the EU-LFS data. Particularly, because of the lack of information on German regions until the mid-2000s (due to legal reasons), we use the European Community Household Panel 1994-2001 to fill those gaps. Furthermore, given that Poland did not disaggregate the sectoral composition of the labour force by region at more than one digit until recently, we obtain the distribution of workers by economic activity (at the two-digit level) and region in Polish manufacturing from the Structural Business Statistics administered by Eurostat.
Our assessment of the effect of robots on job polarisation requires ranking jobs by their average wages (jobs are defined as a combination of a two-digit occupation and a onedigit economic activity, the highest detail level allowed by the EU-LFS micro data). We rank the jobs in the initial year of each of our analyses using the national rankings developed by the European Foundation for the Improvement of Living and Working Conditions (Fernández-Macías et al., 2017). This ranking, based on different data sources of earnings, provides an ordinal classification of jobs by remuneration at two-digit occupation-one-digit sector. Given that there is a significant break in the relevant classifications of occupations (the International Standard Classification of Occupations, ISCO) and industries (the Statistical Classification of Economic Activities in the European Community, NACE) in the end of the 2000s, we reclassify all the jobs in 2015 (using ISCO-08 and NACE Rev. 2) into the previous nomenclature (ISCO-88 and NACE Rev 1.1) relying on a cross-walk constructed on the basis of the double coding used in the European Working Conditions Survey 2010 and 2015 (Eurofound, 2021), performed by Eurofound. There are two additional data sources required for the construction of control variables. The first one is the version of the European Union Capital, Labour, Energy, Material and Service Database (EU KLEMS) available when starting this project (The Conference Board, 2018). The second source is the World Integrated Trade Solution (World Bank, 2021), from which we calculate the increase in the exposure to Chinese trade across European regions.

Methods
In order to disentangle the causal effect of this technology, we exploit the sector-and country-level variation in the deployment of robots. This allows to increase the degrees of freedom of our analysis, in the fashion of , who calculate the exposure to robot adoption by region, assuming that robots' inflows during a time interval follows the distribution of employment (on which we have detailed information) in the initial period of time. Our geographical units of analysis (regions) mainly follow the Nomenclature of Territorial Units for Statistics at the second level (NUTS 2). 2 The right-hand-side variable of interest in our analyses refers to robots. Particularly, in order to put the robot stock in relation with the size of the labour force in the region, we employ the change in the number of robots per region divided by the initial number of workers (in thousands) there. Consequently, we define the increase in robot exposure in region i of a country c as follows where L io denotes the initial number of (thousand) workers in the region at the beginning of the analysed period (time 0); L ikc , the number of workers in industry k in region i at the initial year and L kc0 , the number of workers in industry k in country c at time 0. This definition of robot exposure is identical to the concepts used in , Chiacchio et al. (2018) and Dauth et al. (2021). Unfortunately, the IFR robot database is subject to different measurement problems that have to be accounted for before using the data for econometric analysis. Even in those countries with information for all years and sectors, there is a non-negligible proportion of unspecified stocks and deliveries of robots. Furthermore, in some countries, in the early years there is only information on the total figures of robots, with no reference to the economic sector of deployment. The IFR itself reconstructs the series of robot stocks on the basis of deliveries and a 12-years depreciation assumption (i.e. a robot is fully functional until reaching 12 years of life, when it is withdrawn), with the exception of some countries (e.g. Japan). In principle, the reliability of deliveries' numbers tends to be higher, since the association of robot manufacturers directly controls those inflows, but it can only estimate the stocks. In order to maximise the variability of our main variable of interest (robot adoption), we reconstruct the robot series imputing the unspecified robots under several assumptions. Firstly, we take as given the initial stocks reported in the database and we depreciate them on the basis of the 12-years assumption and the figures observed in the first years of our windows of analysis. Secondly, we rely on robot deliveries because of the reasons stated above. Thirdly, when we impute initial stocks or unspecified deliveries, we base the imputation on the three closest years to the period of interest with specified information.
The main equation of interest is the following where Y i represents the labour market outcome of region i on which we want to assess the impact of robot exposure. Specifically, in our analyses, we explore the effect on employment rates among the working-age population and on employment growth in each tercile of the structure of working population. The latest variable is the result of splitting the employed population into terciles of equal size ranked by the average wage in each job, as mentioned in above. This way, we can determine if robot adoption has a differential impact on different segments of the employment structure, in particular whether it contributes to a process of polarisation, a widespread concern in developed countries nowadays. Although we do not employ any formal measure of polarisation, we look at the estimated coefficients for each tercile and their statistical significance and we compare the estimated coefficients for the three parts of the distribution to assess if these effects significantly differ by tercile. Whilst robotisation is mostly a process that concerns the secondary sector plus mining and quarrying, the analysis of its labour market effects (in terms of overall employment and polarisation) refers to the whole population at work in each region, given that the way in which technology shapes the labour market concerns both directly affected and unaffected sectors. X i comprises a set of control variables, including some covariates referred to the initial period (share of employment in industry in the region, population [thousand people, in logs], share of females, share of people aged 65 or more, share of highly or medium educated population, share of foreign population, average routine-task index [RTI] and average offshorability index). Furthermore, this vector considers the change in ICT capital stock per initial worker in the region of interest and the change in exposure to China net imports per initial worker during the period of interest. The set of covariates are very similar to the ones used by  in their exploration of the effects of Chinese imports shock on American labour markets or the work of Acemoglu and Restrepo (2020) on robot adoption.
The logic of controlling for the initial values of RTI and offshorability is to rule out that the impact of robots conflates with other developments of employment associated to technological changes but alien to robotisation. 3 We compute the region-level change in ICT capital and Chinese net imports from sector-level data, so we have to follow a similar procedure, based on the employment distribution by region and industry, to the one employed for calculating the exposure to robot adoption of each geographical unit. These last two variables are likely to be endogenous, so we include them only to carry out some robustness checks of our main results. It is also worth mentioning that the coverage of capital stocks in the EU KLEMS is far from perfect, sometimes limited to less than 10 countries, depending on the period of analysis. Because of this reason, we often show the results of different econometric specifications with and without the latter control variable. For instance while in some specifications (such as in the analysis of employment using stacked differences) we are able to consider 28 countries, in others, we can just look at 6 (long differences when exploring the effect on employment structure using capital data).
We extend the model in a further specification splitting the sample into two pooled periods (1995-2005 and 2005-2015), including a dummy variable, and adding geographical dummies aiming to capture group-of-countries-specific time trends (Nordic, British Islands, Continental-reference category-, Mediterranean and Eastern Europe). We employ long differences in order to increase the signal-to-noise ratio. Particularly, measurement error is very likely to be present in robot stocks, for the reasons mentioned above. At the same time, within-region variation accounts for most of the variability of the robot stock. In these circumstances, looking at year-by-year differences, the measurement error might have a higher weight than under a long-differences approach. Furthermore, the loss of efficiency due to this setting is minimal here, since we cluster standard errors at the regional level in order to control for serial correlation across geographical units.
Even after controlling for unobserved time-constant heterogeneity (through the first difference) and different sets of covariates, there is room for the existence of regionspecific factors that are potentially correlated with the deployment of robots. This problem would result in inconsistent estimates of β. If, for example robot adoption tends to be faster in the most dynamic regions, the estimated coefficient would be downwards biased. In the absence of natural experiments, we resort to the IV strategy proposed by . These authors build a regional exposure to robotisation based on the penetration of robots in other developedin this case, Europeancountries. Dauth et al. (2021) follow a similar strategy in their study for Germany, trying to exclude border countries and states in the Euro Area, aiming to minimise the existence of common economic shocks among Germany and the countries employed for building the instrument. Chiacchio et al. (2018) use the United Kingdom and Denmark, apart from lagged employment protection legislation, which can raise some endogeneity concerns.
By contrast, in this paper, we use the average change in robot penetration per worker by industry in South Korea, one of the main leaders in robot adoption worldwide, as an instrumental variable. Furthermore, because of the relatively small size of South Korea compared to the global economy, it is unlikely that their sector-level developments could trigger general-equilibrium effects on the countries covered in our analysis. We believe that this instrument might have some advantages over the one used by Chiacchio et al. (2018). They base their IVs on the increase in robot exposure in the United Kingdom and Denmark (which might experience sector-level economic shocks correlated with those in the rest of the continent) and on the Employment Protection Legislation in 1990, which is a dubious exogenous variable, given that labour market rigidity might affect wage performance and that the level of employment protection is likely to be persistent over time.
It is also worth mentioning that, whereas Acemoglu and Restrepo (2020) base their instruments in robot penetration per worker in other countries (hence, putting the number of robots in relation with the country size), this is not the case in the other two papers Dauth et al., 2021). In addition, whilst  and Dauth et al. (2021) are able to replace the share of labour force employed in a certain sector and region in the US and Germany, respectively, by the proportion several decades ago, Chiacchio et al. (2018) do not perform this correction, aiming to discard anticipation effects. We are not able to implement such a correction either, because for most of the countries the first available waves of the EU-LFS correspond to years as late as 1995. To resort to the different censuses does not seem a feasible alternative considering the non-negligible differences among the national industry classifications. But this should not be a large source of concern in the light of the results of  and Dauth et al. (2021). Therefore, we can write down our IV as follows Our IVs, which we implement through 2-stage least squares (2SLS), performs well in most of the cases (F-statistic above 10), with the exception of the second period of analysis (2005)(2006)(2007)(2008)(2009)(2010)(2011)(2012)(2013)(2014)(2015) for some specifications with few countries. Remarkably, whilst in the first period the results are identical irrespective using South Korea or other leading European countries (e.g. Sweden) for the construction of the IV, in the second decade, the strength of IVs dramatically decreases, particularly in the case of other European countries. This is our main motivation for choosing the IV based on South Korea. These problems in the second period of analysis might be related to fact that the crisis affects countries in different ways and the same applies to the adoption of robots (whose stock experiences even decreases in some sectors and countries), so it is very difficult to find an IV that provides enough correlation for all the countries included in the database.

Descriptive statistics
We construct two region-level datasets (of employment and employment by terciles), which share much of their content, but they are not identical. As mentioned above, we carry out a harmonisation of the regions over time and across different databases focussing on the years necessary for performing the econometric analysis (1995, 2005 and 2015), trying to lose as little information as possible. Supplementary Table A1 in the online appendix displays descriptive statistics of the main variables used in the analysis of employment. They are not particularly illustrative, and we show them mainly because of a formal motivation, as the number of available regions changes from 1995-2005 to 2005-2015 because of the EU enlargement. Whereas one can adequately deal with this issue in a regression framework or when evaluating changes, it makes the information embedded in the descriptive statistics of limited usefulness for the reader.
We present in Figure 1 some pairwise correlations between the change in the employment rate and the increase in robot exposure. Focussing on manufacturing only, we can observe a virtual absence of correlation in the first decade and a strong positive correlation over the second time span. The (positive) relationship observed for the 20years time span results from the aggregation of these two heterogeneous sub-periods. Observations are combined by weighting them by the population of each region in 1995. The sample is restricted to the subset of countries for which labour market information is available for the whole period.
In the following tables, we present the most complete specifications, distinguishing between models including and excluding trade and capital (which are both potentially endogenous variables and which, in the case of the latter, reduce the countries available in the database). Models with less control covariates are available upon request.

Econometric results
Effects on employment. Table 3 and 1 display the OLS and 2SLS estimates of the effect of robot adoption on industrial employment. We present estimations for each decade, for the period 1995-2015, and using stacked differences. We also estimate models with our baseline covariates and additionally with exposure to Chinese net imports and the increase in ICT capital. We further consider a set of models without (Table 3) and with groupspecific trends (Table 1). In principle, it is a sign of robustness that the estimates do not change much when including these controls. We further argue that including these dummies controls to some extent for the existence of different trends in employment creation across Europe that might confound the impact of robots. However, there is some debate about the convenience of including such dummies in some econometric settings as they might capture dynamic effects of the regressor of interest, particularly, in differencesin-differences designs (Wolfers, 2006). In most of our cases, the F-statistic of our IV is above the common threshold of 10, indicating its relevance, although in the case of stacked differences its strength decreases. They are particularly weaker in the second period of analysis (2005)(2006)(2007)(2008)(2009)(2010)(2011)(2012)(2013)(2014)(2015), where the patterns of adoption across countries differ more than in the pre-crisis period and, therefore, are overall less synchronised with South Korea. Although it performs better than other instruments proposed in the literature in this context based on the potential replaceability of labour at the sector-level in the US several decades ago (see, also, Bekthiar et al., (2021) for a detailed assessment), this should be a limitation of our analysis that we should bear in mind. Table 2 presents the results without controlling for group-specific trends. In the first period, we find a negative association between robotisation and employment in the case of  Notes: Standard errors clustered at the regional level between parentheses. *** significant at 1% level; ** significant at 5%; * significant at 10%. All models include a constant, as et of start-of-period variables (share of employment in industry, population, share of females, share of population aged 65 and over, share of population with medium or high education, average RTI index and the average offshorability index and regional dummies (when possible, Nordic, British, Mediterranean and East, with Continental as reference). The specification using stacked differences includes a dummy for the period 2005-2015. Observations are weighted by total regional population at the beginning of each period. Countries in I, II, VII and VIII: the 2SLS estimates. The coefficients reflect the percent effect of an increase of a robot per 1000 workers in the initial period in the region on the share of regional working-age population employed in industry. For instance an effect of À0.009 means that one additional robot per 1000 workers leads to a reduction of 0.9% in the share of working-age individuals employed in industry in the region. Unsurprisingly, the 2SLS estimates under not very strong IV is quite imprecise ranging from barely 0 to À0.02, which means quite much uncertainty on the impact of robot adoption. For the OLS estimator coefficients are positive and even smaller but not significant. In the second period, the results of both the OLS and 2SLS estimates are positive and of similar magnitudes, but they are only significant for the OLS estimator. Looking at the full period using stacked differences, we find a positive and significant association for both OLS and 2SLS and independent of the trade and capital controls. Using a long-difference estimator we find significant (but small) negative effects in the case of the 2SLS with capital and trade controls. All other estimates using a long-difference approach are not significant, small and with varying sign. In most cases, controlling for trade and capital does not make much of a difference.
Results accounting for group-specific trends are reported in Table 2. In the first period, we can see by comparing the first two columns of Table 1 and 3 that when we include the control covariates (specifically, region-specific trends, that account for the possibility that the path of employment growth is different across different groups of countries), the Notes: Standard errors clustered at the regional level between parentheses. *** significant at 1% level; ** significant at 5%; * significant at 10%. All models include a constant, as et of start-of-period variables (share of employment in industry, population, share of females, share of population aged 65 and over, share of population with medium or high education, average RTI index and the average offshorability index), the change in the net exposure to Chinese imports and regional dummies (when possible, Nordic, British, Mediterranean and East, with Continental as reference). Observations are weighted by total regional population at the beginning of the period.  Notes: Standard errors clustered at the regional level between parentheses. *** significant at 1% level; ** significant at 5%; * significant at 10%. All models include a constant, as et of start-of-period variables (share of employment in industry, population, share of females, share of population aged 65 and over, share of population with medium or high education, average RTI index and the average offshorability index). The specification using stacked differences includes a dummy for the period 2005-2015. Observations are weighted by total regional population at the beginning of each period. coefficients of the 2SLS estimate remain negative and significant whilst the coefficients for the OLS estimates now also become negative and, when controlling for trade and capital, statistically significant. 4 In the second period, results become positive and significant in most cases and the coefficients are relatively large for the 2SLS estimator, compared to the other estimates. It appears that the strength of our IV decreases in the second period of analysis. This can be related to the heterogeneity of economic performance of countries during the Great Recession and subsequent recovery, also affecting robot adoption. 5 We have checked how sensitive the results in the second period might be with regard to the selection of countries. Looking at the same countries as in the specification including capital and even only at a subset of six countries (Finland, France, Germany, Italy, Spain and Sweden), correlations are positive in both OLS and IV estimates and in many cases statistically significant.
Looking at the entire period the results are very similar to the case without covariates described in Table 3. For the stacked differences, the coefficients are positive, but only significant for the 2SLS estimator. For long differences, the coefficients are very small and negative, but only significant when trade and capital controls are included.
There are discrepancies between OLS and 2SLS particularly in the first period (1995)(1996)(1997)(1998)(1999)(2000)(2001)(2002)(2003)(2004)(2005), when we do not control for region-specific time trends. The larger OLS estimates would suggest that those regions more dynamic in terms of employment would be deploying more robots. These differences become much smaller when we include time trends. The estimated coefficients go in the same direction, the differences are not large and often are not statistically different from zero. This could reflect just the lower precision of 2SLS rather than a situation where the most dynamic regions in terms of employment deployed more robots in terms of employment in the first period and the most dynamic ones would present such behaviour later.
We present the results for the whole economy in supplementary Table A2 in the Annex. They are essentially in line with the ones presented in Table 3 and 2: the effects tend to be negative in the first period, whilst in the second period they tend to be positive. When looking at the entire period, stacked differences suggest a relatively large positive effect, whilst the long-difference approach is inconclusive. In this respect, our results from the first period resemble those of Dauth et al. (2021), who find negative employment effects only in manufacturing sectors, not in the whole economy, whilst our results from the second period and from the stacked differences approach for the entire period appear to confirm the positive association between robotisation and employment found in Klenert et al. (2021).
We estimate several alternative specifications, with similar results. In the first place, using stacked differences, we excluded Eastern European countries. Secondly, we have estimated our model including only a balanced panel of regions. In the third place, we look at the impact of employment in the same subset of countries as Chiacchio et al. (2018), which mainly comprises countries with higher robot densities. Finally, we have estimated our main regression for the period 1995-2007. These results, not shown for brevity, are qualitatively similar to the ones presented here.
Effects on the employment structure. In this section we analyse the impact of robot adoption on different parts of the distribution of employment across occupations. Assessing the effect of this technology on different segments of the labour market is a way to check whether robots have contributed to the phenomenon of job polarisation, that is decline of employment in mid-skilled or mid-paid occupations relative to the highest and lowest. We adopt a simplified approach for characterising polarisation, comparing the effect of the variable of interest on employment in the middle occupational tercile relative to the extremes. If this difference is very small, we consider that there is no polarisation effect. The dependent variable is employment in each occupational tercile. The terciles are obtained by ranking jobs, a crossing of two-digit ISCO and NACE classifications, by their average wages and then assigning them to three groups of equal size in terms of initial employment.
The results for 1995-2005 and 2005-2015 using stacked differences and long differences are given in Tables 1 and 4, supplementary table A3 and A4, respectively. Each table contains both the OLS and the 2SLS estimates. In all cases, we have included our main baseline controls and group-specific trends and we present another specification including capital and trade.
Regarding the first decade of the analysis (Table 3), our results point to a significant negative effect on the middle tercile, which could be viewed as evidence for polarisation Notes: Standard errors clustered at the regional level between parentheses. *** significant at 1% level; ** significant at 5%; * significant at 10%. All models include a constant, as et of start-of-period variables (share of employment in industry, population, share of females, share of population aged 65 and over, share of population with medium or high education, average RTI index and the average offshorability index), the change in the net exposure to Chinese imports and regional dummies (when possible, Nordic, British, Mediterranean and East, with Continental as reference). Observations are weighted by total regional population at the beginning of the period. Countries (results for the bottom and top tercile are not significant). However, equality tests indicates that the effect on the middle tercile is not significantly different from the effect in the bottom and the top terciles. We find that the estimated coefficient for the middle tercile is significantly different from the coefficient for the top tercile. However, it is fairly similar to the bottom coefficient. Rather than indicating polarisation, this finding suggests upgrading (a relative expansion of the top tercile). Nevertheless, we should interpret this result with some caution, given the outcome of the test and the limited statistical power of our analysis. During the second period (Table 4), results for the OLS estimator are significant for all terciles when looking at the specification without capital, but we observe no evidence of polarisation. When capital and trade controls are included, the increase in employment is significantly different from zero and positive in the bottom and the top tercile, which could be interpreted as a sign of job polarisation. However, our statistical tests based do not allow rejecting that the estimated coefficient for the middle quintile differ from neither the bottom nor the top one. In the case of the long differences (Supplementary Table A3), none of the estimated coefficients are statistically significant.
Finally, Supplementary Table A4 displays the results of the model using stacked differences. In this case, the results are somewhat puzzling. In the case of the OLS, the coefficients are not statistically significant. When looking at the 2SLS estimates, we observe a pattern that differs from the one shown in Tables 3 and 4. The coefficients are positive, but the coefficient of the middle tercile is larger than the coefficients of the bottom and top terciles, which would in fact suggest a reversal of job polarisation. These results should be interpreted with caution, since the strength of our instrument decreases substantially when using stacked differences whilst the magnitudes of the coefficient do not. The reasons for such a large deviation of the OLS from the IV estimates remain unclear.
For brevity, we do not present the analyses that exclude regional dummies. These results, available upon request from the authors, are roughly in line with the results presented in Supplementary Table A4. There are some signs of polarising effects in the period 1995-2005, with comparatively less negative impact of employment in the bottom and the top of the employment distribution than in the middle. However, the differences between the change in the middle tercile and at least one of the extremes is not always significant. In the case of the period 2005-2015, where the results suggest a certain polarisation (but accompanied by employment growth), the results excluding dummies support less of a polarising pattern than the results presented in Tables 3 and 4,  Supplemetary Tables A3 and A4.
From our assessment on the impact of robotisation on the labour market structure, we do not think that any robust conclusion can be drawn.

Discussion
The results presented in the previous section are to some extent puzzling, as the association between robots and employment appears to be predominantly negative in the first half of the observed period and largely positive in the second half. This leads to mostly inconclusive and sometimes positive estimates when pooling the observations to include both periods. In this section, we discuss possible reasons behind this outcome and compare our results to the literature.
There are several possible explanations for the discrepancy between the two periods. In the first place, there might be some sort of unobserved heterogeneity that we are not able to control for (e.g. the differential geographic impact of the economic crisis of 2008).
Secondly, the differences in the results between periods might be influenced by the effect of robotisation on productivity. Whilst the potential negative effect of robot adoption on employment is immediate and likely to be constant over time, the potential positive impact associated with productivity gains manifests itself later (Acemoglu and Restrepo, 2018), which could explain the positive effect in the second period. In this fashion, Jungmittag and Pesole (2019) find a positive effect of robot adoption on productivity during a similar time window, which increases during the second decade (2005)(2006)(2007)(2008)(2009)(2010)(2011)(2012)(2013)(2014)(2015). These authors argue that the higher rise in productivity in the latter period might be due to the existence of nonlinearities, such as a minimum level of robot stock being required for driving such a positive impact. Restrepo (2018, 2020) suggest that the effect of automation on productivity might be larger in the long term because capital accumulation stabilises the price of capital. In their framework, wages, and the wage bill, can rise in the long term, but employment diminishes. It is possible that the distribution of costs and benefits could be different in the context of the models of industrial relations in Europe. There is qualitative evidence on these sorts of processes of adaptation to technology (Eurofound, 2021;Grande et al., 2021). It is possible that the larger output results in an increase of the labour demand that counteracts the employment losses. Autor and Salomons (2018) suggests that this could have been the case during the last decades. Although this labour-augmenting effect of technology would apply to a lower extent to new technologies, we should bear in mind that the robots we consider here are a quite mature and far from revolutionary one. Furthermore, it is possible that the most affected categories of workers have been already pushed out of the labour market. Finally,  and Koch et al. (2021) provide some evidence that robot adoption has a positive effect on employment at the firm level but might have a negative one on competitors. It is possible that the diffusion of the technology over time might reduce the negative size of robot adoption.
Thirdly, the effect of robotisation is generally small in magnitude (which can be seen from the estimation coefficients) and the quality of the data simply might not allow it to be isolated, a problem that is further aggravated when looking at the installation of robots in regions instead of sectors. As shown by Fernández-Macías et al. (2021), robots are highly concentrated in specific manufacturing sectorsattributing them to different regions might further complicate identification of the effect among other factors that impact employment in the region.
Fourthly, we include countries in the second period for which no data was available in the first period. In particular, Eastern European countries exhibited a strong growth in robotisation (see Fernández-Macías et al., 2021) as well as an expanding economy after 2005.
Fifthly, we should not rule out that these results are due to another sort of heterogeneity: we consider a 20-year period, during which robotic technology may have significantly advanced, so it is conceivable that more recent robots have a higher impact on productivity or increased complementarity to the existing labour force than older models, even against the backdrop of a fraction of replaceable workers already having been displaced.
Regarding other literature on this topic, we are able to roughly reproduce the results obtained by Chiacchio et al. (2018) for the first period and a similar set of countries, particularly when looking at employment in mining, quarrying, manufacturing, electricity, water supply and construction. In the case of the whole economy, the effects are weaker but still present. Our findings also resemble those of , who find a negative correlation for a period of analysis from 1993 to 2007.
In the second period (2005)(2006)(2007)(2008)(2009)(2010)(2011)(2012)(2013)(2014)(2015), the impact of robots on employment is positive for the majority of model specifications. It also is mostly positive when looking at both periods using a stacked differences approach. This suggests positive effects of robotisation on employment similar to those found by Dahlin, (2019); Klenert et al. (2021). However, they use a different approach, which is based on the sector-level penetration of robots in Europe and the United States, respectively, not on the regional distribution of robots. Graetz and Michaels (2018), who also use an approach based on the classification of robots by sector, do not find a correlation between robotisation and total employment for the period 1993-2007. In sum, this striking heterogeneity of results suggests that the effects of industrial robots on employment in the period studied are relatively small and depend to a great extent on the model specifications and the selection of countries and years.
We should also acknowledge the limitations imposed by the number of regions included in the analysis, which determine the degrees of freedom and the statistical power of our econometric tests. Although we make a considerable effort in enlarging the sample as much as possible, it is undeniable that this issue imposes an important constraint and, in many cases, it is difficult to assess the extent to which the lack of significance is driven by the low number of observations or by the actual absence of non-negligible effects. In relation to this issue, we should also keep in mind that the relevance of the instrument decreases substantially during the second decade, which is probably related to the larger unobserved heterogeneity during this turbulent economic period.

Conclusions
Nowadays, few topics generate as much interest in both public and academic circles as the impact of technology on the labour market. Within this context, our paper aims to explore the effect of industrial robots, a very specific technology, on employment and labour market polarisation in regional European labour markets. Our main contribution is twofold: first, we replicate and extend previous research on this topic in several ways, covering additional countries and years as well as exploring a broader range of instruments and specifications. Secondly, we analyse the impact of robot adoption on the occupational structure, to check whether this technology contributes to job polarisation.
Our results suggest that the effects of robots on regional labour markets are generally not particularly robust, as they appear to differ over time: our estimates for the period 1995-2005 suggest a negative association between robotisation and employment in Europe, whilst in the second period (2005)(2006)(2007)(2008)(2009)(2010)(2011)(2012)(2013)(2014)(2015), this association becomes positive.
Regarding the impact on the employment structure, there is some weak evidence that robotisation might have polarised the labour market to some extent in the period 1995-2005. However, this effect is not observed in the second period. Interestingly, for certain model specifications, job polarisation appears to be reversed when looking at the whole period ; that is we find a more pronounced positive effect of robotisation on the middle tercile. For all periods analysed, however, the statistical significance greatly depends on the model specifications and therefore these results should be interpreted with additional caution.
Previous studies that analyse the employment effect of robots in local labour markets with similar methods only use data up to 2007 . It is therefore not surprising that we confirm their finding of a negative association between robots and employment in the first period. Studies that instead use an approach based on a classification of robots by sector, instead of region, appear to find a positive effect of robots on aggregate employment  or no effect at all (Graetz and Michaels, 2018). In contrast to Graetz and Michaels (2018); Klenert et al. (2021) also analyse the period until 2015 and find an overall positive correlation between robots and employment.
The analysis in this paper is inevitably subject to several limitations. Firstly, despite having different hypotheses regarding the differential pattern of the effects of robots on employment over time, we cannot prove their validity econometrically. We leave this challenge to future work. Secondly, we cannot rule out factors such as the economic turbulences due to the Great Recession in the second decade confounding the results, as the extent and depth of the crisis differed significantly across Europe. Thirdly, a caveat this work shares with other recent works using the IFR data is the low precision of the underlying data on robots, compounded in our case by the limited number of regions, which reduces the precision of the estimates. Fourthly, the fact that industrial robots are highly concentrated in certain sectors and countriesfor instance German car manufacturing accounted for 27% of all robot stocks in Europe in 2016might distort the results to some extent . Fourthly, industrial robots are only one specific type of automation technology, which happens to be sufficiently welldocumented for performing econometric analysis. Hence, our results cannot be generalised to other automation technologies, such as service robots or AI, which might well have a large and significant impact on employment.
Our findings have important implications for policymakers. The variability and the small magnitude of our results show that the assertion that robots have a negative effect on employment is far from granted. In this respect, calls for regulating or taxing industrial robots might not be justified. 6 Such policies might in fact reduce productivity, as industrial robots have been found to be positively correlated with productivity (Graetz and Michaels, 2018;Jungmittag and Pesole, 2019).
One can speculate about the relationship between robot adoption and industrial relations on the basis of the recent literature on this topic. Firstly, it suggests the presence of some mechanisms for dealing with the disruptive effects of robots that can be found in several niches of employment that suggest the consequences on industrial relations would not be large, as long as they as long as they do not seem to involve plant closures or massive layoffs or leaving a number of middle-age workers with bad job prospects helpless. For instance Dauth et al. (2021) finds that robot adoption tends to favour the promotion of incumbent manufacturing workers to higher-level tasks in Germany. This process reduces job opportunities in manufacturing for young people, who tend to stay longer in education to compensate it. The study of Lech (2020) for the US indicates that early retirement means another possible exit for older employees in the most affected sectors. Secondly, there is qualitative evidence on the role of social dialogue on mitigating the costs of introducing disruptive automation technology (minimising dismissals in a container terminal) and the absence of effects on industrial relations (in Airbus) (Eurofound, 2021;Grande et al., 2021).