Robotization, internal migration and rural depopulation in Austria

Bekhtiar, Karim Working Paper Robotization, internal migration and rural depopulation in Austria IHS Working Paper, No. 41 Provided in Cooperation with: Institute for Advanced Studies (IHS), Vienna Suggested Citation: Bekhtiar, Karim (2022) : Robotization, internal migration and rural depopulation in Austria, IHS Working Paper, No. 41, Institut für Höhere Studien - Institute for Advanced Studies (IHS), Vienna This Version is available at: https://hdl.handle.net/10419/260595 Standard-Nutzungsbedingungen: Die Dokumente auf EconStor dürfen zu eigenen wissenschaftlichen Zwecken und zum Privatgebrauch gespeichert und kopiert werden. Sie dürfen die Dokumente nicht für öffentliche oder kommerzielle Zwecke vervielfältigen, öffentlich ausstellen, öffentlich zugänglich machen, vertreiben oder anderweitig nutzen. Sofern die Verfasser die Dokumente unter Open-Content-Lizenzen (insbesondere CC-Lizenzen) zur Verfügung gestellt haben sollten, gelten abweichend von diesen Nutzungsbedingungen die in der dort genannten Lizenz gewährten Nutzungsrechte. Terms of use: Documents in EconStor may be saved and copied for your personal and scholarly purposes. You are not to copy documents for public or commercial purposes, to exhibit the documents publicly, to make them publicly available on the internet, or to distribute or otherwise use the documents in public. If the documents have been made available under an Open Content Licence (especially Creative Commons Licences), you may exercise further usage rights as specified in the indicated licence. https://creativecommons.org/licenses/by/4.0/ IHS Working Paper 41 June 2022 Robotization, Internal Migration and Rural Depopulation in Austria Karim Bekhtiar Author(s) Karim Bekhtiar Editor(s) Robert M. Kunst Title Robotization, Internal Migration and Rural Depopulation in Austria Institut für Höhere Studien - Institute for Advanced Studies (IHS) Josefstädter Straße 39, A-1080 Wien T +43 1 59991-0 F +43 1 59991-555 www.ihs.ac.at ZVR: 066207973 Funder(s) Anniversary Fund, Austrian National Bank (OeNB - grant number 18462), Austrian Economic Association (NOeG) NOeG-Dissertation Fellowship 2022 License „Robotization, Internal Migration and Rural Depopulation in Austria“ by Karim Bekhtiar is licensed under the Creative Commons: Attribution 4.0 License (http://creativecommons.org/licenses/by/4.0/) All contents are without guarantee. Any liability of the contributors of the IHS from the content of this work is excluded. All IHS Working Papers are available online: https://irihs.ihs.ac.at/view/ihs_series/ser=5Fihswps.html This paper is available for download without charge at: https://irihs.ihs.ac.at/id/eprint/6194/ Robotization, Internal Migration and Rural Depopulation in Austria Karim Bekhtiar1 June 8, 2022 Abstract Internal migration flows from rural to urban areas have greatly contributed to population declines in many rural areas across both Europe and the US. At the same time there is mounting evidence for a tight connection between internal migration and shifts in labor demand, with the latter being heavily affected by the rise of automation technologies. Therefore this paper analyzes the effects industrial robotization has had on manufacturing employment and internal migration in Austria during the period 2003-2016, specifically focusing on rural-to-urban migration flows. The results show that robotization has caused significant declines in manufacturing employment to which populations reacted by increased out-migration. This migratory response takes the form of rural-to-urban migration, thereby contributing to population declines in many rural areas in Austria. These rural-to-urban movements are primarily driven by young and medium/low skilled individuals, i.e. those groups that bear the strongest shock incidence. JEL classification: J21, J23, J61, R23 Keywords: Employment, internal migration, robots, rural depopulation 1Institute for Advanced Studies (IHS), Vienna and Johanes Kepler University (JKU), Linz Contact: [email protected]. I want to thank Rudolf Winter-Ebmer for exceptional support and guidance. Furthermore I want to thank Rene B¨oheim, Wolfgang Frimmel, Michael Irlacher, Michael Pfaffermayer, Markus Walzl and various seminar participants at the Central European University in Vienna, the RWI workshop for causal inference with spatial data in Essen, the University of Innsbruck, the Johanes Kepler University Linz and the Institute for Advanced Studies (IHS) for useful comments. I further want to thank Statistics Austria for providing the migration flow data. The research presented in this paper was funded by the anniversary fund of the Austrian National Bank (OeNB - grant number 18462) and has received additional financial support from the Austrian Economic Association (NOeG) via the NOeG-Dissertation Fellowship 2022. 1 Introduction: Over the last decades population declines in remote rural areas have become a persistent feature of demographic change in both Europe and the US. As young and highly educated individuals increasingly migrate towards the cities, declining rural regions are left with lasting declines in human capital (Bjerke and Mellander,2017) and economic performance (Dax and Fischer,2018), the disappearance of many private and public services (Rickardsson,2021) and drastic shifts in the age structure (Johnson, Field, and Poston Jr.,2015). As a consequence rural depopulation contributes to increased geographic inequality and fosters political polarization, as ”the left behind” (Wuthnow,2018) show increasingly strong support for populist political movements. Correspondingly, the support of declining rural regions contributed strongly to the victory of Donald Trump in the 2016 US elections (Scala and Johnson,2017,Wuthnow, 2018), Brexit (Lee, Morris, and Kemeny,2018) and the electoral success of far-right populist parties in several European countries (see for example Franz, Fratzscher, and Kritikos,2018,Rickardsson,2021 or Kenny and Luca,2021). Despite the detrimental impact of rural depopulation on economic, social and political cohesion in Europe and the US, most of the research concerned with the causes of rural-to-urban migration is focused on low and middle income countries.1A notable exception to this is the recent work of Johnson and Lichter (2019) who document large and persistent rural-to-urban migration flows for the US. With regards to the causes of rural-to-urban migration they argue for a close connection to declines in manufacturing employment in rural areas. This mirrors a well established notion in the economic literature that internal migration flows play a crucial role in the reaction to regional labor 1See for example Zhao (1999), Brueckner and Lall (2015), Lagakos, Mobarak, and Waugh (2018), Peri and Sasahara (2019) or Lagakos (2020). 1 demand shocks.2Since labor demand in the manufacturing industries has experienced growing pressure in recent decades through the rise of automation technologies3, there thus appears to be scope for a connection between rural depopulation and automation. Therefore this paper analyses the effects industrial robotization has had on internal migration and, more specifically, rural-to-urban migration in Austria during the period 2003–2016. For this I use detailed data on municipality-to-municipality migration flows. Following Acemoglu and Restrepo (2020) and Dauth et al. (2021) changes in robotization are predicted as a shift-share variable, using regional industry structures and industry-level data on robot densification from the International Federation of Robotics (IFR). To isolate the causal effect of robotization on internal migration and rural depopulation, predicted robot exposure is instrumented with a shift-share-type instrumental variable, which is constructed from industry level robotization trends in other high income countries. As is shown in Borusyak, Hull, and Jaravel (2022) leveraging plausibly exogenous variation in robotization shocks in other high income countries isolates the component of robot adoption that is driven by exogenous advances in technological possibilities. Applying this identification strategy to the Austrian data confirms a robust negative effect of robotization on manufacturing employment and a positive effect on out-migration flows, indicating that robotization has had displacement effects in highly exposed local labor markets, which in turn led to migratory responses of affected work- 2Prominent examples of this literature are Blanchard and Katz (1992), Bound and Holzer (2000), Cadena and Kovak (2016), Huttunen, Møen, and Salvanes (2018), Foote, Grosz, and Stevens (2019), Greenland, Lopresti, and McHenry (2019) or Notowidigdo (2020). For employment shocks caused by industrial robots, Faber, Sarto, and Tabellini (2021) have recently shown that robotization caused declines in the working age-population of particularly affected US local labor markets. 3See for example Autor, Levy, and Murnane (2003), Goos and Manning (2007), Autor, Katz, and Kearney (2008), Autor and Dorn (2013), Goos, Manning, and Salomons (2014), or Acemoglu and Restrepo (2020) among many others. 2 ers. Decomposing these migration flows by the type of origin and destination region (urban or rural) reveals that robotization led to out-migration in affected rural areas, with the majority of this out-migration taking the form of rural-to-urban migration flows, thereby contributing to rural depopulation. This effect on rural-to-urban migration flows is primarily driven by the demographic sub-groups whose employment prospects are most heavily affected by the robotization shock, namely by young and medium to low skilled individuals. This paper relates to the extensive literature on the effects of industrial robots on labor market outcomes, as well as the literature on migratory responses to local labor demand shocks. It contributes to this literature by (i) showing that robotization shocks are mitigated by out-migration in a similar fashion as other large scale labor demand shocks and (ii) connecting these migratory responses to a highly relevant demographic trend in recent decades – rural depopulation. To the best of my knowledge, this paper is the first to present causal evidence on a connection between shifts in labor demand and rural depopulation. The rest of this paper is structured as follows: Section 2presents a descriptive overview over population trends in rural Austria. Section 3presents the used data sources, while Section 4discusses the empirical approach and the identification strategy. Section 5presents the main results of the analysis and explores the robustness of these results. Lastly, Section 6offers a brief discussion and concludes. 2 Rural Depopulation in Austria To illustrate the close connection between out-migration and general population trends in rural regions, Figure 1compares the change in overall population counts (panel A) and the migration balance (panel B) of all Austrian municipalities between 3 2001 and 2016.4While urban areas generally showed increases in population counts and net in-migration, a large fraction of rural municipalities experienced population losses through out-migration. These declining rural municipalities tend to be in more remote areas of Austria, as rural regions in closer proximity to urban centers also experienced population growth. While some rural regions thus appear to benefit from positive population spillovers from nearby urban centers (Veneri and Ruiz,2016), around 46% of all rural municipalities show a negative migration balance for the period 2001 to 2016 (Table 1, panel A). By 2016 those declining rural municipalities on average lost about 6.4% of their 2001 population (panel B, column 4). These population losses in declining rural areas are largely driven by out-migration, which on average accounts for a population loss of around −3.53%. The remainder of the population loss is explained by the fact that individuals who leave a municipality are typically younger, than those who stay behind, leading to older societies and declines in the birth balance. This highlights that rural out-migration not only directly decreases population counts in declining rural areas, but also has an indirect negative effect through the acceleration of natural decline (Johnson, Field, and Poston Jr.,2015). Distinguishing between internal- and external-migration flows in panel B of Table 1reveals that in the absence of migration from other countries (i.e. in the absence of external migration) the average net outflows in rural communities with declining populations would be much stronger with an average internal net outflow of around 4Urban areas and rural areas are classified according to the urban-rural-classification from the Austrian statistical agency Statistics Austria. This classification consists of three broad categories of municipalities: urban centers, regional centers and rural areas, each consisting of several subcategories. It is graphically depicted in Figure A1 in the Appendix. For this paper, I consider municipalities classified as ’urban centers’ (large, medium or small) as urban, while all remaining municipalities (including regional centers) are classified as rural. All results presented in this paper are robust to this choice. 4 Figure 1: Population trends (2001-2016) by municipality type (a) Population Change (2001-2016) (b) Net Migration (2001-2016) Note: Municipalities are classified according to the urban-rural classification from the Austrian Statistical Agency (Statistics Austria - see Figure A1 in the Appendix). Large urban centers (according to the urban-rural classification) are indicated by name. All remaining urban centers (medium and small) are indicated by lighter colors. Rural areas and regional centers are depicted as a single category. Population data is from the decennial census (2001) and the registry based labor market statistics (2016). Data on net migration-flows is taken from the migration statistics. All data sources are available from Statistics Austria and are described in more detail in Section 3. 5 local population (measured in the initial year of each panel period to avoid endogenous contamination).9The inclusion of the composition of the local population is motivated by concerns that regions with different demographic structure are very likely to be on different trends regarding population changes and migration-flows. Furthermore it has been shown in recent work by Acemoglu and Restrepo (2022) that the structure of the workforce (particularly the age composition) has a direct impact on robotization trends.10 Therefore this very detailed set of demographic variables is included to control for this simultaneous impact of the demographic structure on robotization and migration trends. In the second step, controls that aim at capturing regional heterogeneity are included. These controls include the start-of-period logarithm of the gross regional product and the regional unemployment rate (to control for differences in economic performance) and the start-of-period share of the population living in urban areas (to control for different population trends depending on the degree of urbanization). The next set of covariates controls for other types of labor demand shocks. For this I include shift-share variables for changes in import- and export-exposure from 9The composition of the local population is included in 68 age-sex-education-nationality cells, where each cell indicates the size of the respective group in the overall population in 1000 individuals. The 68 demographic cells are defined by 5 age groups (ages 0-14, 15-34, 35-49, 50-64 and 65 and above), 2 gender groups (male, female), 4 educational groups (highest level of education completed is either compulsory schooling, apprenticeship, high-school or university) and 2 nationalities (Austrian or foreign citizen). 10Acemoglu and Restrepo (2022) document the fact that jobs that are automatable by industrial robots are predominantly performed by middle aged workers, since these are the workers which mainly perform routine manual tasks. As aging reduces labor supply from middle aged workers (and thereby increases their wage rate) the relative price of robots vis-a-vis those workers drops. This makes robotization more profitable and thus leads to stronger robot adoption in sectors which rely more heavily on middle aged workers. 12 China and the former Eastern Block,11 as well as ICT-capital intensity. As is laid out in detail by Adao, Koles´ar, and Morales (2019) other types of labor demand shocks that can be expressed as shift-share type variables have a mechanical correlation with ∆Robotsr,t, since they are constructed from similar exposure shares. Therefore these control variables have to be included to control for other large scale labor demand shocks originating in international trade or other forms of automation technologies. Data on import- and export exposure comes from the UN-Comtrade database, while data on ICT-intensity is taken from the EU-Klems database.12 By a similar logic I also control for shocks to labor supply originating in migration from foreign citizens. Following Card (2001), the migrant share in a region is a strong predictor for migration inflows. To the extent that this migrant share correlates with the industry exposure shares used in equation 1to predict regional robot exposure, migration based labor supply shocks are also mechanically correlated with predicted robot exposure (see Adao, Koles´ar, and Morales,2019). Therefore I include changes in the migrant population, differentiated by 4 educational groups. Labor supply related controls are however only included in employment regressions, because migration flows of non-Austrian citizens are themselves a component of population changes and migration 11This follows the approach of Dauth, Findeisen, and Suedekum (2014) who showed for Germany that trade with countries of the former Eastern Block is more relevant for the German case. Therefore the shift-share variables for import- and export-exposure are computed as changes in import- and export-exposure from China, Bulgaria, Czech Republic, Hungary, Poland, Romania, Slovakia, Slovenia, and the succession states of the former USSR - Russian Federation, Belarus, Estonia, Latvia, Lithuania, Moldova, Ukraine, Azerbaijan, Georgia, Kazakhstan, Kyrgyzstan, Tajikistan, Turkmenistan and Uzbekistan. 12The Comtrade data, which is only available at the commodity level, has been crosswalked to the ISIC-Rev. 4/NACE-Rev. 2 classification using the comtradr-package in R. The EU-Klems data comes from the September 2017 release. 13 flows (i.e. the dependent variables in all regressions except employment regressions). Lastly I control for the regional start-of-period industry structure, to check for the possibility that regions with different industry structure are on different trends, both in robotization as well as in changes in employment or migration flows. This is done in two different ways. Firstly, the share of manufacturing employment is included as additional control. Secondly, instead of the manufacturing share, a more detailed set of industry structure controls are included.13 Fixed Effects: To control for unobserved shocks that influenced all regions equally, all estimations include a set of period fixed effects. As the observational period ranges from 2003 to 2016, these period fixed effects are of particular importance, as they aim to control for confounding effects of the Great Recession. Therefore the two panel periods are defined such that they correspond to a pre-crisis period (2003-2009) and a postcrisis/recovery period (2009-2016).14 However, as is emphasized in Borusyak, Hull, and Jaravel (2022), these period fixed effects require some adjustments. Robot adoption is strongly concentrated within the manufacturing sectors. Therefore most industries outside of manufacturing experienced zero robotization. In equation 1this means that the regional sum of all exposure shares of sectors with non-zero robot adoption is gen- 13For the detailed industry composition controls I follow Dauth et al. (2021) and include the initial period employment shares of sub-industries of manufacturing (production of food products, consumer goods, industrial goods and capital goods), as well as industries outside of manufacturing (construction, personal services, business services and the public sector). 14As can be seen in Figure A2 in the Appendix robotization in Austria was rather unaffected by the Great Recession, as the increase in robotization continued rather smoothly. This somewhat mitigates concerns about a confounding influence of the Great Recession. However the period fixed effects are included to more thoroughly control for this. 14 erally smaller than 1, such that Pi Empi,r Empr<1. As is explained in detail in Borusyak, Hull, and Jaravel (2022), conventional period fixed effects do not properly isolate within period variation in shift-share applications with incomplete exposure shares. To correct this, they recommend to interact the period fixed effects with the regional sum of the incomplete exposure shares, as only in this case the fixed effects fully absorb between period variation. Therefore the period fixed effects τtin equation 2refer to the interaction of conventional period dummies with the regional sum of the incomplete exposure shares. In OLS estimations the incomplete shares used for the computation of predicted robot exposure in equation 1are used, while in 2SLS regressions the lagged exposure shares used for the computation of the instrument (from equation 3– see section 4.1) are used. Additionally all estimations contain a set of region fixed effects ρrto control for unobserved regional heterogeneity. Commuting Zones: To control for the possibility of spatial spillovers of local robotization shocks, I use so-called commuting zones (instead of the municipalities themselves) as unit of observation. Clearly a local shock to a plant in municipality kdoes not only influence employment (and outcomes related to employment) in the same municipality. Rather it is to be expected that employment in neighboring municipalities will react as well, simply because some workers who worked in the same plant, and thus are directly affected by the shock, commuted there from neighboring areas. Also a number of studies have shown that local shocks influence employment in other plants in the same local labor market via agglomeration effects (see for example Gathmann, Helm, and Sch¨onberg 2018 or Helm 2019). To control for these spillover effects more aggregated local labor markets are often used in the literature to examine the consequences of local shocks. 15 The underlying idea is that aggregated regional units more closely correspond to local labor markets, and therefore spillovers caused by commuting patterns are less of an issue than when using the municipalities themselves. While some studies use existing geographical units, which are usually defined as administrative areas, districts or states, to approximate local labor markets, I use commuting zones as they are a data driven way to define local labor markets based on the strength of their commuting ties. Using commuting zones appears to be less arbitrary than using predefined administrative areas, and has a straightforward appeal as they are specifically designed to contain a larger fraction of commuters within their borders. The construction of these commuting zones strictly follows the methodology used for US commuting zones described in Tolbert and Sizer (1996) and Dorn (2009) (for details see Online Appendix C).15 Standard Errors: Throughout the analysis I report two types of standard errors, namely conventional heteroskedasticity robust standard errors, as well as the exposure robust standard errors from Adao, Koles´ar, and Morales (2019) (henceforth referred to as AKM-standard errors). As is outlined in detail in Adao, Koles´ar, and Morales (2019) conventional standard errors might be unreliable in shift-share settings, as the regression residuals are likely to be correlated across (potentially distant) regions with similar exposure shares. To account for this possibility I report both sets of standard errors.16 15The US commuting zones estimated by Tolbert and Sizer (1996) are widely used in studies on labor market shocks in the US. See for example Autor, Dorn, and Hanson (2013,2015,2021), Autor and Dorn (2013), Acemoglu et al. (2016), or Acemoglu and Restrepo (2020) among others. 16Adao, Koles´ar, and Morales (2019) show in their paper that the AKM-standard errors might be downward biased and thus overreject if the number of industries used to construct the shift-share variable is too small. Since the IFR-data only includes 26 different industries, which can be used for the computation of ∆Robotsit, this is a potential concern in this setting. To address this concern, I 16 4.1 Identification Strategy: One major reason for endogeneity concerns in equation 2is that the adoption of robots might be correlated with unobserved regional demand shocks which simultaneously influence employment trends or internal migration decisions. For example, negative shocks to the domestic demand for goods produced by industry imight reduce that industry’s demand for industrial robots. Such demand shocks could be related to changes in employment or migration flows in areas where industry iis a relevant part of the local economy. In such a scenario the estimate for γin equation 2would no longer isolate the effect of industrial robotization but would additionally reflect effects arising from the unobserved demand shock. Another source for endogeneity concerns relates to the construction of the predicted robotization measure in equation 1. Here the industry level change in robotization is assigned to any region rpurely via the regional structure of employment. This implicitly assumes that all firms in a given industry iare equally likely to adopt robots. Any violation of this assumption leads to a measurement error in the explanatory variable which, to the degree that it is systematically related to unobserved regional characteristics, would lead to a bias in the estimate for γ. Consider for example the presence of regional agglomeration effects that incentivise high performing firms to settle in a certain region. If high performing firms are also more likely to adopt industrial robots (as recent findings in Bonfiglioli et al.,2020 and Koch, Manuylov, and Smolka,2021 suggest) the predicted robotization measure in equation 1would have a measurement error that is systematically related to this unobserved agglomeration effect. To address these concerns I follow Acemoglu and Restrepo (2020) and Dauth et al. apply a modification to the computation of these standard errors that results in more conservative estimates. The details and performance of this modification are discussed in Online Appendix D. 17 (2021) and construct an instrumental variable that leverages exogenous variation in robot adoption from other high-income countries. Since industry level robotization trends in other high-income countries are unrelated to unobserved regional characteristics in any Austrian region (like regional demand shocks or agglomeration economies), this approach isolates changes in the supply of robots which is driven by advances in the technological frontier. Similarly to predicted robot exposure in equation 1, this instrumental variable is constructed as a shift-share variable, where industry level robotization changes in other high-income countries are interacted with regional exposure shares. ∆RobotsIV r,t =X i Empi,r,t−15 Empr,t−15 ×∆RobotsOtherCountries i,t Empi,t−15 (3) To further remove the instrumental variable in equation 3from the predicted robot exposure measure in equation 1the exposure shares used to construct the instrument are lagged by 15-years. As has been shown in recent work by Adao, Koles´ar, and Morales (2019) and Borusyak, Hull, and Jaravel (2022), the validity of this instrumental variable hinges on the exogeneity of the industry level robotization shocks occurring in other high-income countries. The underlying identifying assumption thus is that industry level robotization trends in other high-income countries ∆RobotsOtherCountries i,t are quasi-randomly assigned with respect to unobserved regional characteristics in Austria. In the examples described above this means that the robotization trends in other high income countries must not have a direct impact on region specific demand shocks in Austria or the location decisions of robotizing Austrian firms. As is shown in Borusyak, Hull, and Jaravel (2022) this exogeneity of the robotization shocks is both necessary and sufficient for the instrumental variable to be valid. Hence the regional exposure shares (i.e. the 18 lagged industry structure) are allowed to be endogenous.17 To construct these robotization shocks occurring in other high income countries I use industry level robotization changes in Canada, Denmark, Finland, France, Italy, Mexico, Norway, Spain, Sweden, the United Kingdom and the United States.18 While the exogeneity of the robotization shocks, which essentially mirrors a standard exclusion restriction, cannot be tested directly, Borusyak, Hull, and Jaravel (2022) propose several plausibility tests. These tests aim to assess the plausibility of quasi-random shock assignment by assessing whether the robotization shocks themselves and the constructed instrument are balanced (i.e. not systematically related) to pre-determined characteristics in Austria. The following two subsections present these balance tests on the industry and on the regional level. Industry Level Balance Tests To assess the quasi-random assignment of the industry level shocks used to construct the instrument, Table 2presents industry level balance tests. These tests are conducted by estimating the following industry level regression: YAustria i,t =β∆RobotsOtherCountries i,t +τt+i,t (4) 17In a related paper Goldsmith-Pinkham, Paul, and Swift (2020) argue that the exogeneity of the exposure shares is also a sufficient condition for the validity of the instrumental variable. Borusyak, Hull, and Jaravel (2022) however show that the orthogonality of the shocks is both sufficient and necessary and that in the Goldsmith-Pinkham, Paul, and Swift (2020) setting of exogenous regional exposure shares, shock exogeneity is implicitly fulfilled due to the exogenous (i.e. quasi random) assignment of the regional exposure shares. 18Canada, Mexico and the United States are not available as separate countries in the IFR-data, but are rather aggregated to a single region (North America). 19 where start-of-period values of some observed industry level characteristic in Austria YAustria i,t are regressed on the industry level robotization changes in other countries ∆RobotsOtherCountries i,t which are also used in equation 3to construct the regional level instrumental variable. To isolate within-period variation of the robotization shocks, the estimations control for period fixed effects. The regression in equation 4is estimated separately for each of the start-of-period characteristics of industries in Austria YAustria i,t . These balance variables aim to test for the balance of the shocks with respect to the industry level composition of the workforce, as well as other industry characteristics related to productivity, capital intensity and the average wage rate. Data on industry level workforce characteristics has been calculated from the ASSD, while all remaining indicators have been calculated from EU-Klems data.19 The results of the industry level balance tests in Table 2show that industry level changes in robot adoption in the countries used to construct the instrument are significantly correlated with the age composition of the workforce in Austria, while the estimates for all other balance variables are insignificant. The measure for the age composition used in this estimation is computed analogously to Acemoglu and Restrepo (2022) who use the ratio of old to middle aged workers to show that aging of the workforce (which would be equivalent to an increase in this ratio) leads to increases in robot adoption. The significant and negative estimate of −0.854 for this balance test indicates that greater robot adoption in the other high income countries predicts a stronger reliance on middle aged workers in Austrian industries.20 19Because the ASSD does not contain any information on the skill level of workers, no industry level balance variables related to the skill composition of the workforce are available. Similarly EU-Klems only includes such information up to 2005. Therefore balance tests related to the skill composition cannot be performed at the industry level and are postponed to regional level tests. 20This correlation likely stems from (i) the age structure having an influence on robotization changes (as shown by Acemoglu and Restrepo,2022) and (ii) industry level age structures being correlated 20 Table 2: Industry level balance tests for instrumental variable: Balance variable Coef. SE (1) (2) Start-of-period ratio of old workers to middle aged workers -0.854 (0.422)** Start-of-period share of blue collar workers 1.467 (1.659) Start-of-period labor productivity: 0.108 (0.204) Start-of-period capital/labor ratio: -0.515 (1.034) Start-of-period ICT-capital/capital stock -0.032 (0.055) Start-of-period log(avg. hourly real wage) 0.018 (0.019) Industries: 26 Time Periods: 2 Industry-Period Shocks: 52 Notes: *<0.10, ** <0.05, *** <0.01. This Table shows industry level regressions of the respective balance variables on the industry level robotization shocks ∆RobotsOtherCountries i,t used in equation 3to construct the instrumental variable. The industry level robotization shocks are summed up over all countries and are then normalized to have zero-mean and unit variance. The ratio of old workers to middle aged workers is constructed by dividing industry level employment of workers aged 50 or older, by employment of workers age 35 to 49. Industry level data on employment by age and worker type (blue collar) is taken from the ASSD data, while all remaining industry level balance variables are taken from the EU-KLEMS September 2017 Release (July 2018 Update). All regressions control for period fixed effects and are weighted by industry size. Heteroskedasticity robust standard errors are reported in brackets. In sum these industry level balance tests in Table 2show that the shocks used to compute the instrument are reasonably balanced, with the exception of the age composition of the workforce. This suggests that it is crucial to control for workforce characteristics in the regional level estimations, as the exogeneity of the robotization shocks (and thus the exogeneity of the entire instrument) is likely only fulfilled when conditioning on demographic characteristics of the workforce. Regional Level Balance Tests Table 3presents estimation results for regional level balance tests. For this I follow the recommendations in Borusyak, Hull, and Jaravel (2022) and Goldsmith-Pinkham, across countries. Table A5 in the Appendix shows that this is indeed the case, as the industry level age structure of the workforce in Austria is significantly correlated with the age structures in countries used to construct the instrument. 21 incumbent manufacturing workers. The reduction in new hirings however strongly dominates the reduction in separations, leading to the overall negative effect on employment in the manufacturing sector. This result suggests that the burden of the robotization shock primarily falls on individuals seeking to enter new employment, rather than displacing incumbent workers. This result is very similar to results for the German case in Dauth et al. (2021) who also find that the majority of the disemployment effect in manufacturing falls on non-incumbent workers. Table A2 in the Appendix presents estimation results for the effect on manufacturing employment decomposed by age-groups. Here the results show that the majority of the shock incidence (41% of the effect on manufacturing employment and 46% of the effect on blue-collar employment) falls on younger workers below the age of 35. The robotization shock thus particularly hampers the employment prospect of young workers, a group that is known in the literature to be more geographically mobile in response to labor demand shocks (see for example Bound and Holzer,2000). Internal Migration & Rural Depopulation: To examine whether these disruptions in labor demand caused by industrial robots have led to increased out-migration, Table 5presents estimations of the effect of robotization on net out-migration rates. For any period tthat spans the years j= 1, ..., J this measure is constructed as: PJ j=1 Net Outflowj Populationj=1 =PJ j=1(Outflowj−Inflowj) Populationj=1 (6) Hence net out-migration rates are calculated by subtracting migration-inflows from migration-outflows, and summing up over all years that make up the panel period. The measure is then normalized by the initial year working age population to arrive at a relative measure of net out-migration flows. 28 Table 5: Robotization and Internal Migration (2003-2016) Dependent Variable: Net-outmigration-rate ×100 (1) (2) (4) (5) (6) OLS: ∆ Robots 0.22 0.203 0.261 0.229 0.155 (0.116)* (0.13) (0.15)* (0.15) (0.165) [0.049]*** [0.046]*** [0.042]*** [0.042]*** [0.043]*** 2SLS: ∆ Robots 0.833 1.173 1.078 1.105 1.055 (0.353)** (0.459)** (0.356)*** (0.362)*** (0.323)*** [0.064]*** [0.073]*** [0.066]*** [0.067]*** [0.071]*** First Stage Results: 0.011 0.011 0.011 0.011 0.011 (0.001)*** (0.001)*** (0.001)*** (0.001)*** (0.001)*** [0.0003]*** [0.0003]*** [0.0003]*** [0.0003]*** [0.0002]*** First Stage F-Statistic: 37.09 34.58 27.94 27.58 29.77 Period Fixed Effects x x x x x Region Fixed Effects x x x x x Demographic Controls x x x x x Regional Characteristics x x x x Labor Demand Shifts x x x Manufacturing Share x Detailed Industry Structure x Commuting Zones 158 158 158 158 158 Periods 2 2 2 2 2 Observations 316 316 316 316 316 Notes: *<0.10, ** <0.05, *** <0.01. Conventional robust standard errors are shown in round brackets and shiftshare robust standard errors from Adao, Koles´ar, and Morales (2019) are shown in square brackets. Units of observation are 158 clustered commuting zones (for details see Online Appendix C). All specifications include a set of region and period fixed effects, whereby the period fixed effects are interacted with the sum of exposure shares used to construct the explanatory variable (OLS) or the instrument (2SLS). Demographic controls include the start-of-period structure of the local population in 68 age-gender-education-nationality cells. Regional characteristics control for the start-of-period logarithm of the gross regional product and the unemployment rate, as well as the start-of-period degree of urbanization. Labor Demand Shifts include changes in import- and export-exposure and ICT-intensity. The detailed industry structure controls include start-of-period employment shares of several sub-industries of manufacturing (production of food products, consumer goods, industrial goods and capital goods), as well as industries outside of manufacturing (construction, personal services and business services) and the public sector. All regressions are weighted by start-of-period working age population. 29 The estimation results for the migratory response in Table 5show that robotization has led to an increase in net out-migration rates during 2003-2016. Here the full specification in column 5 suggests that one more industrial robot per 1000 workers leads to net out-migration flows of around 1.1% of the start-of-period working age population. This effect is robust over all specifications and statistically highly significant in both standard error definitions. Comparing the results for the OLS and 2SLS estimations shows that the 2SLS point estimates are drastically larger than the OLS estimates. This picture is consistent with recent findings in Borusyak, Dix-Carneiro, and Kovak (2022) who show that the OLS- estimates from migration regressions are (at times severely) biased towards zero when shocks between origin and destination regions are correlated. The sizable difference between the OLS and 2SLS estimates suggests that the instrumentation strategy is able to successfully address this issue, as the 2SLS estimation results in a rather large estimated effect.22 While the results in Table 5confirm that robotization shocks led to out-migration in a similar fashion as is firmly established for other types of labor demand shocks, these results remain silent about the direction of these internal migration flows. To lay a specific focus on the question whether robotization causes migration flows directed from rural to urban areas, and thereby contributes to rural depopulation, I use the fact that the data on net out-migration rates used in Table 5contains detailed information on the municipality of origin, as well as the destination. As any commuting zone may consist of both urban and rural areas (see Online Appendix C), the net outflow from 22Even if the instrumentation strategy would not be able to fully address this problem, the estimated effect in column 6 of Table 5would be a lower bound of the true migration response, as Borusyak, Dix-Carneiro, and Kovak (2022) show that a correlation between shocks in origin and destination regions would always result in an attenuation of the estimate towards zero. 30 any commuting zone can be decomposed into the respective contributions of rural and urban areas: PJ j=1 Net Outflowj Populationj=1 =PJ j=1 Net OutflowRural j Populationj=1 +PJ j=1 Net OutflowUrban j Populationj=1 (7) Using the available information on the destination type (urban, rural or abroad), the net outflows from rural areas can be further decomposed by destination: PJ j=1 Net OutflowRural j Populationj=1 =PJ j=1 Net OutflowRural→Urban j Populationj=1 + PJ j=1 Net OutflowRural→Rural j Populationj=1 + PJ j=1 Net OutflowRural→Abroad j Populationj=1                    Internal        External (8) Hence the net out-migration rate from all rural municipalities in any commuting zone is decomposed into flows directed towards urban or rural areas (internal migration) and flows with other countries (external migration).23 Table 6applies this decomposition to the net out-migration rates from rural areas. Here column 1 shows the effect of industrial robots on all rural net outflows. This effect (0.953) is on a very similar magnitude as when net outflows from both rural and urban 23While the migration flow data contains detailed information on the municipality of origin, it does not contain information on the exact municipality of destination. Rather the type of destination is provided (as defined in the urban-rural-classification from Statistics Austria in Appendix A1). While this does not allow a detailed reconstruction of the destination municipality, it allows for a distinction between rural and urban destinations. 31 Table 6: Robotization and net out-migration in Rural Areas (2003-2016) Total External Internal All Rural to Urban Rural to Rural (1) (2) (3) (4) (5) OLS: ∆ Robots 0.141 -0.001 0.142 0.057 0.085 (0.164) (0.024) (0.181) (0.09) (0.103) [0.037]*** [0.005] [0.041]*** [0.02]*** [0.025]*** 2SLS: ∆ Robots 0.953 -0.067 1.02 0.565 0.455 (0.307)*** (0.048) (0.342)*** (0.162)*** (0.202)** [0.061]*** [0.009]*** [0.069]*** [0.031]*** [0.042]*** First-Stage F: 29.77 29.77 29.77 29.77 29.77 Period Fixed Effects x x x x x Region Fixed Effects x x x x x Demographic Controls x x x x x Regional Characteristics x x x x x Labor Demand Shifts x x x x x Detailed Industry Structure x x x x x Commuting Zones 158 158 158 158 158 Periods 2 2 2 2 2 Observations 316 316 316 316 316 Notes: *<0.10, ** <0.05, *** <0.01. Conventional robust standard errors are shown in round brackets, and shift-share robust standard errors from Adao, Koles´ar, and Morales (2019) are shown in square brackets. Units of observation are 158 clustered commuting zones (for details see Online Appendix C). All specifications include a set of region and period fixed effects, whereby the period fixed effects are interacted with the sum of exposure shares used to construct the explanatory variable (OLS) or the instrument (2SLS). Demographic controls include the start-of- period structure of the local population in 68 age-gender-education-nationality cells. Regional characteristics control for the start-of-period logarithm of the gross regional product and the unemployment rate, as well as the start-of-period degree of urbanization. Labor Demand Shifts include changes in import- and export-exposure and ICT-intensity. The detailed industry structure controls include start-of-period employment shares of several sub-industries of manufacturing (production of food products, consumer goods, industrial goods and capital goods), as well as industries outside of manufacturing (construction, personal services and business services) and the public sector. All regressions are weighted by start-of-period working age population. areas are considered (1.055 - Table 5, column 6). Decomposing this effect into the part explained by external migration (column 2) and internal migration (column 3) makes clear that increases in net out-migration flows are exclusively driven by increases in internal net out-migration. Decomposing these internal migration flows into rural-to- urban and rural-to-rural flows in columns 4 and 5 of Table 6reveals that a large part of this effect stems from rural-to-urban migration. While one more additional robot per 1000 workers increases internal out-migration rates from rural areas by around 1%, approximately 0.57% of this increase are accounted for by outflows that are directed 32 towards urban areas. This effect for rural-to-urban net migration rates is precisely estimated and significant at the 1%-level in both standard error definitions. This coefficient in column 4 of Table 6provides direct evidence that robotization contributes to population declines in rural areas by specifically increasing rural-to-urban internal migration flows. Column 5 of Table 6also provides evidence that robotization has an increasing effect on rural-ro-rural migration flows. Here the estimation suggests a positive effect of 0.455. This effect of robotization on rural-to-rural migration flows is somewhat smaller than the effect on rural-to-urban migration flows, highlighting that the increase in net outmigration rates from rural areas strongly operates through rural-to-urban migration.24 Since the dependent variables used in the estimations shown in Table 6are computed by subtracting in-migration-flows from out-migration-flows (to arrive at the desired net out-migration measure in equation 6) it is interesting whether the increase in rural- to-urban net out-migration stems from an increase in out-migration flows, or rather a decrease in in-migration-flows. To answer this question Table 7presents separate estimations on those two components of net out-migration rates. Comparing columns 2 and 3 of Table 7shows that the increase in net out-migration rates is exclusively driven by an increase in out-migration (column 2), while the estimate for the effect of robotization on in-migration is small and statistically insignificant in both standard error definitions. 24The classification into rural and urban areas used in Table 6is performed according to the urbanrural-classification from Statistik Austria (see Appendix Figure A1). This classification distinguishes between three broad categories (urban centers, regional centers and rural areas), each being comprised of several subcategories. In Table 6the intermediate category ’regional centers’ is classified as rural area. To check whether this choice affects the results, Table A4 in the Appendix presents estimates, where ’regional centers’ are included in urban areas. The estimates in Table A4 show that the results are unaffected by this choice. 33 Table 7: In-migration and out-migration in Rural Areas (2003-2016): Net Out-Migration Out-migration In-migration (1) (2) (3) OLS: ∆ Robots 0.057 0.03 -0.027 (0.09) (0.099) (0.062) [0.02]*** [0.021] [0.015]* 2SLS: ∆ Robots 0.565 0.586 0.021 (0.162)*** (0.169)*** (0.109) [0.031]*** [0.034]*** [0.021] First-Stage F: 29.77 29.77 29.77 Period Fixed Effects x x x Region Fixed Effects x x x Demographic Controls x x x Regional Characteristics x x x Labor Demand Shifts x x x Detailed Industry Structure x x x Commuting Zones 158 158 158 Periods 2 2 2 Observations 316 316 316 Notes: *<0.10, ** <0.05, *** <0.01. Conventional robust standard errors are shown in round brackets, and shift-share robust standard errors from Adao, Koles´ar, and Morales (2019) are shown in square brackets. Units of observation are 158 clustered commuting zones (for details see Online Appendix C). All specifications include a set of region and period fixed effects, whereby the period fixed effects are interacted with the sum of exposure shares used to construct the explanatory variable (OLS) or the instrument (2SLS). Demographic controls include the start-of- period structure of the local population in 68 age-gender-education-nationality cells. Regional characteristics control for the start-of-period logarithm of the gross regional product and the unemployment rate, as well as the start-of-period degree of urbanization. Labor Demand Shifts include changes in import- and export-exposure and ICT-intensity. The detailed industry structure controls include start-of-period employment shares of several sub-industries of manufacturing (production of food products, consumer goods, industrial goods and capital goods), as well as industries outside of manufacturing (construction, personal services and business services) and the public sector. All regressions are weighted by start-of-period working age population. Taken together the results in tables 6and 7clearly show that robotization has increased migration flows from rural to urban areas. As this specific type of internal migration flow greatly contributes to population declines in many rural areas, these results show that robotization based labor demand disruptions have contributed to rural depopulation in Austria between 2003 and 2016. To benchmark the magnitude of this effect Panel B of Figure A3 in the Appendix presents a counterfactual calculation, where robotization is held constant at its 2003 level. This Figure shows that between 2003 and 2016 rural areas in Austria lost around 34 3.62% of their 2003 working age population through rural-to-urban net outflows. In the absence of robotization this number drops to around 2.78%. Increases in robotization thus explain around one fourth of all rural-to-urban migration flows during the period 2003 to 2016. 5.1 Robustness Checks: Pre-Trend Tests To test whether the results in tables 4to 7are indeed driven by exogenous changes in the robotization shocks used to construct the instrument rather than by pre-trends in the outcome variables, Table 8presents pre-trend tests. For these tests changes in the outcome variables between different pre-periods during 1991 and 2003 are regressed on the instrument for 2003 to 2016. The pre-trend test for the log-change in manufacturing employment is presented in panel A of Table 8. Because the migration flow data used for the computation of net out-migration rates used in tables 5to 7are only available starting in 2002, these pre-trend tests can not be performed using migration-flow data. To assess the presence of pre-trends in internal migration, I therefore rely on an alternative measure for migration responses. For this I use the log-change in working age population counts (panel B) and the log-change in the rural working age population (panel C).25 Because 25In the literature on migratory responses to local labor demand shocks the log-change in working age population counts is frequently used as primary measure for migration flows (especially when more detailed data on migration in- and outflows is not available). Table A3 in the Appendix shows that using the log-change in working age population counts to approximate migration responses yields similar results as when using net out-migration rates (as in Table 5). Consistent with previous results, these estimations suggest that robotization leads to significant decreases in the size of the working age population, both overall (Table A3, panel A) and when focusing only on rural areas (panel B). 35 these population counts are taken from the Austrian decennial census, pre-period data points are only available for the years 1991 and 2001. Therefore the pre-trend tests for log-changes in population related variables in columns 1 to 3 of Table 8contain linearly interpolated population counts for the years 1997 and 2003. To be sure that the results are not driven by the linear interpolation of the population data, column 4 presents an additional pre-trend test which only relies on data from census years. All pre-trend tests are conducted in two variations. Firstly I regress pre-period changes on the instrumental variable, controlling only for period fixed effects. Secondly I include all available control variables. Here it is important to note that, since the pre-period changes do not vary across the two panel periods, regional fixed effects can not be included in pre-trend tests including a full set of controls. The estimation results for the pre-trend tests in Table 8(panel B) suggest the presence of slightly negative pre-trends in the log-change of the working age population in the pre-periods 1991-1997 (column 1) and 1991-2001 (column 4). These pre-trends are however only significant using the conventional heteroskedasticity robust standard errors, while the shift-share-robust AKM-standard errors suggest that these pre-trends are statistically insignificant. Significant pre-trends in the log-change of the working age population are potentially concerning, as they indicate that regions with higher values for the instrument were on a negative population trend prior to the treatment. Conditioning on all available controls however reduces the estimates in size and renders these pre-trends statistically insignificant. All remaining pre-trend tests for log-changes in manufacturing employment (panel A) and log-changes in the rural working age population (panel C) result in insignificant estimates, indicating parallel pre-trends. In sum the pre-trend tests in Table 8are reassuring, as they indicate parallel pretrends in all pre-periods. Only in the log-change of the working age population in panel B of Table 8do these parallel pre-trends hold only conditionally on the available control 36 Table 8: Pre-trend tests 1991-1997 1997-2003 1991-2003 1991-2001 (1) (2) (3) (4) Panel A: Dependent Variable: Pre-Trend for ∆ log(manufacturing employment) Control only for period FE -0.0038 0.0031 -0.0007 (0.0074) (0.0094) (0.0132) [0.1288] [0.1247] [0.2534] Full controls 0.0054 -0.0001 0.0053 (0.0092) (0.0084) (0.0103) [0.0103] [0.0085] [0.0094] Panel B: Dependent Variable: Pre-Trend for ∆ log(working-age population) Control only for period FE -0.0027 -0.001 -0.0037 -0.0043 (0.0012)** (0.0014) (0.0022) (0.0019)** [0.004] [0.0118] [0.0156] [0.0067] Full controls -0.0007 0 -0.0007 -0.001 (0.0014) (0.0014) (0.0027) (0.0023) [0.0011] [0.001] [0.002] [0.0018] Panel C: Dependent Variable: Pre-Trend for ∆ log(rural working-age population) Control only for period FE -0.0025 -0.0017 -0.0042 -0.004 (0.002) (0.0019) (0.004) (0.0033) [0.027] [0.0296] [0.0566] [0.044] Full controls 0.0006 0.0011 0.0018 0.0012 (0.0013) (0.0013) (0.0026) (0.0022) [0.001] [0.001] [0.0019] [0.0017] Commuting Zones 158 158 158 158 Periods 2 2 2 2 Observations 316 316 316 316 Notes: *<0.10, ** <0.05, *** <0.01. Conventional robust standard errors are shown in round brackets, and shift-share robust standard errors from Adao, Koles´ar, and Morales (2019) are shown in square brackets. Units of observation are 158 clustered commuting zones (for details see Online Appendix C). Period fixed effects are interacted with the sum of exposure shares used to construct the explanatory variable (OLS) or the instrument (2SLS). Since the pre-trend variables do not vary between time periods, regional fixed effects are omitted in regressions including a full set of controls. All regressions are weighted by start-of-period working age population. 37 Together these two groups account for around 78% of all migratory responses to the robotization shock. 6 Conclusion It has been long established in the economic literature that internal migration plays a crucial role in the recovery of local labor markets after large scale shocks to labor demand. While this mechanism is well known in the context of general labor demand shocks, the impact of industrial robotization on internal migration flows only recently received some attention. Also the question where internal migrants move after a shock remained largely unstudied. This question is however of particular relevance as internal migration flows are a major contributing factor to population declines in many rural areas in both Europe and the US. This phenomenon, which is known as rural depopulation, poses a great challenge for many rural areas, and also for society as a whole, as it is closely connected to increases in geographical inequality and social and political polarization. In this paper I explore the connection between changes in labor demand which are caused by the rise of industrial robotization, internal migration and rural depopulation in Austria during the period 2003 to 2016. The results of the analysis show that industrial robotization has had a substantial negative impact on manufacturing employment, and increased out-migration in local labor markets most exposed to the robotization shock. Laying a specific focus on rural areas reveals that these internal out-migration flows in rural areas are primarily directed towards urban areas, thereby contributing to the decline of many rural regions. In sum the estimations suggest that rural-to-urban migration flows which are specifically caused by industrial robotization explain roughly one fourth of all rural-to-urban movements between 2003 and 2016. Exploring heteroge- 44 neous effects by population subgroups further shows that these rural-to-urban migration flows are primarily driven by those individuals that bear the strongest incidence of the robotization shock, namely young and medium- to low-skilled individuals. 45 References Acemoglu, Daron, David H. Autor, David Dorn, Gordon H. Hanson, and Brendan Price. 2016. “Import Competition and the Great U.S. Employment Sag of the 2000s.” Journal of Labor Economics 34 (S1):S141–S198. Acemoglu, Daron and Pascual Restrepo. 2020. “Robots and Jobs: Evidence from US Labor Markets.” Journal of Political Economy 128 (6):2188–2244. ———. 2022. “Demographics and Automation.” Review of Economic Studies 89 (1):1– 44. Adao, Rodrigo, Michal Koles´ar, and Eduardo Morales. 2019. “Shift-Share Designs: Theory and Inference.” The Quarterly Journal of Economics 134 (4):1949–2010. Autor, David and Mark G. Duggan. 2003. “The rise in the disability rolls and the decline in unemployment.” The Quarterly Journal of Economics 118 (1):157–206. Autor, David H. and David Dorn. 2013. “The Growth of Low-Skill Service Jobs and the Polarization of the US Labor Market.” American Economic Review 103 (5):1553– 1597. Autor, David H., David Dorn, and Gordon H. Hanson. 2013. “The China Syndrome: Local Labor Market Effects of Import Competition in the United States.” American Economic Review 103 (6):2121–2168. ———. 2015. “Untangling Trade and Technology: Evidence from Local Labour Markets.” The Economic Journal 125 (584):621–646. ———. 2021. “On the Persistence of the China Shock.” NBER Working Paper 29401 . 46 Autor, David H., Lawrence F. Katz, and Melissa S. Kearney. 2008. “Trends in U.S. Wage Inequality: Revising the Revisionists.” The Review of Economics and Statistics 90 (2):300–323. Autor, David H., Frank Levy, and Richard J. Murnane. 2003. “The Skill Content of Recent Technological Change: An Empirical Exploration.” The Quarterly Journal of Economics 118 (4):1279–1333. Bartik, Timothy J. 1991. “Who benefits from state and local economic development policies?” . Bjerke, Lina and Charlotta Mellander. 2017. “Moving home again? Never! The locational choices of graduates in Sweden.” The Annals of Regional Science 59 (3):707– 729. Blanchard, Olivier J. and Lawrence F. Katz. 1992. “Regional Evolutions.” Brookings Papers on Economic Activity 23 (1):1–76. Bonfiglioli, Alessandra, Rosario Crin`o, Harald Fadinger, and Gino Gancia. 2020. “Robot imports and firm-level outcomes.” . Borusyak, Kirill, Rafael Dix-Carneiro, and Brian K Kovak. 2022. “Understanding Migration Responses to Local Shocks.” . Borusyak, Kirill, Peter Hull, and Xavier Jaravel. 2022. “Quasi-Experimental Shift-Share Research Designs.” Review of Economic Studies 89 (1):181–213. Bound, John and Harry J. Holzer. 2000. “Demand Shifts, Population Adjustments, and Labor Market Outcomes during the 1980s.” Journal of Labor Economics 18 (1):20– 54. 47 Brueckner, Jan K and Somik V Lall. 2015. “Cities in developing countries: fueled by rural–urban migration, lacking in tenure security, and short of affordable housing.” Handbook of regional and urban economics 5:1399–1455. Cadena, Brian C. and Brian K. Kovak. 2016. “Immigrants Equilibrate Local Labor Markets: Evidence from the Great Recession.” American Economic Journal: Applied Economics 8 (1):257–290. Card, David. 2001. “Immigrant inflows, native outflows, and the local labor market impacts of higher immigration.” Journal of Labor Economics 19 (1):22–64. Dauth, Wolfgang, Sebastian Findeisen, and Jens Suedekum. 2014. “The Rise of the East and the Far East: German Labor Markets and Trade Integration.” Journal of the European Economic Association 12 (6):1643–1675. Dauth, Wolfgang, Sebastian Findeisen, Jens Suedekum, and Nicole Woessner. 2021. “The adjustment of labor markets to robots.” Journal of the European Economic Association 19 (6):3104–3153. Dax, Thomas and Michael Fischer. 2018. “An Alternative Policy Approach to Rural Development in Regions Facing Population Decline.” European Planning Studies 26 (2):297–315. Dorn, David. 2009. Essays on inequality, spatial interaction, and the demand for skills. Ph.D. thesis. Faber, Marius, Andr´es Sarto, and Marco Tabellini. 2021. “Local Shocks and Internal Migration: The Disparate Effects of Robots and Chinese Imports in the US.” . Foote, Andrew, Michel Grosz, and Ann Stevens. 2019. “Locate Your Nearest Exit: Mass Layoffs and Local Labor Market Response.” ILR Review 72 (1):101–126. 48 Franz, Christian, Marcel Fratzscher, and Alexander S Kritikos. 2018. “German rightwing party AfD finds more support in rural areas with aging populations.” DIW Weekly Report 8 (7/8):69–79. Gathmann, Christina, Ines Helm, and Uta Sch¨onberg. 2018. “Spillover Effects of Mass Layoffs.” Journal of the European Economic Association 18 (1):427–468. Goldsmith-Pinkham, Isaac Paul, Sorkin, and Henry Swift. 2020. “Bartik instruments: What, When, Why, and How.” American Economic Review 110 (8):2586–2624. Goos, Maarten and Alan Manning. 2007. “Lousy and Lovely Jobs: The Rising Polarization of Work in Britain.” The Review of Economics and Statistics 89 (1):118–133. Goos, Maarten, Alan Manning, and Anna Salomons. 2014. “Explaining Job Polarization: Routine-Biased Technological Change and Offshoring.” American Economic Review 104 (8):2509–2526. Graetz, Georg and Guy Michaels. 2018. “Robots at Work.” The Review of Economics and Statistics 100 (5):753–768. Greenland, Andrew, John Lopresti, and Peter McHenry. 2019. “Import Competition and Internal Migration.” Review of Economics and Statistics 101 (1):44–59. Greenstone, Michael, Alexandre Mas, and Hoai-Luu Nguyen. 2020. “Do Credit Market Shocks Affect the Real Economy? Quasi-experimental Evidence from the Great Recession and ”Normal” Economic Times.” American Economic Journal: Economic Policy 12 (1). Helm, Ines. 2019. “National Industry Trade Shocks, Local Labour Markets, and Agglomeration Spillovers.” The Review of Economic Studies 87 (3):1399–1431. 49 Hummels, David, Rasmus J˜ A¸rgensen, Jakob Munch, and Chong Xiang. 2014. “The Wage Effects of Offshoring: Evidence from Danish Matched Worker-Firm Data.” American Economic Review 104 (6). Huttunen, Kristiina, Jarle Møen, and Kjell G. Salvanes. 2018. “Job Loss and Regional Mobility.” Journal of Labor Economics 36 (2):479–509. Johnson, Kenneth M., Layton M. Field, and Dudley L. Poston Jr. 2015. “More Deaths Than Births: Subnational Natural Decrease in Europe and the United States.” Population and Developement Review 41 (4):651–680. Johnson, Kenneth M. and Daniel T. Lichter. 2019. “Rural Depopulation: Growth and Decline Processes over the Past Century.” Rural Sociology 84 (1):3–27. Kenny, Michael and Davide Luca. 2021. “The urban-rural polarisation of political disenchantment: an investigation of social and political attitudes in 30 European countries.” Cambridge Journal of Regions, Economy and Society 14 (3):565–582. Klump, Rainer, Anne Jurkat, and Florian Schneider. 2021. “Tracking the rise of robots: A survey of the IFR database and its applications.” . Koch, Michael, Ilya Manuylov, and Marcel Smolka. 2021. “Robots and firms.” The Economic Journal 131 (638):2553–2584. Lagakos, David. 2020. “Urban-rural gaps in the developing world: Does internal migration offer opportunities?” Journal of Economic perspectives 34 (3):174–92. Lagakos, David, Ahmed Mushfiq Mobarak, and Michael E Waugh. 2018. “The welfare effects of encouraging rural-urban migration.” NBER Working Paper 24193 . Lee, Neil, Katy Morris, and Thomas Kemeny. 2018. “Immobility and the Brexit vote.” Cambridge Journal of Regions, Economy and Society 11 (1):143–163. 50 Notowidigdo, Matthew J. 2020. “The Incidence of Local Labor Demand Shocks.” Journal of Labor Economics 38 (3):687–725. Peri, Giovanni and Akira Sasahara. 2019. “The impact of global warming on rural- Urban migrations: Evidence from global big data.” NBER Working Paper 25728 . Rickardsson, Jonna. 2021. “The urban–rural divide in radical right populist support: the role of residents characteristics, urbanization trends and public service supply.” The Annals of Regional Science 67 (1):211–242. Scala, Dante J. and Kenneth M. Johnson. 2017. “Political polarization along the ruralurban continuum? The geography of the presidential vote, 2000–2016.” The ANNALS of the American Academy of Political and Social Science 672 (1):162–184. Tolbert, Charles M. and Molly Sizer. 1996. “US Commuting Zones and Labor Market Areas: A 1990 Update.” Economic Research Service Staff Paper 9614 . Veneri, Paolo and Vicente Ruiz. 2016. “Urban-To-Rural Population Growth Linkages: Evidence from OECD Tl3 Regions.” Journal of Regional Science 56 (1):3–24. Wuthnow, Robert. 2018. The Left Behind: Decline and Rage in Rural America. Princeton University Press. Zhao, Yaohui. 1999. “Leaving the countryside: rural-to-urban migration decisions in China.” American Economic Review 89 (2):281–286. Zweim¨uller, Josef, Rudolf Winter-Ebmer, Rafael Lalive, Andreas Kuhn, Jean-Philippe Wuellrich, Oliver Ruf, and Simon B¨uchi. 2009. “Austrian Social Security Database.” Working Paper No. 0903, NRN: The Austrian Center for Labor Economics and the Analysis of the Welfare State. . 51 Appendix A Additional Descriptives & Results: Figure A1: Urban-Rural Classification from Statistics Austria (2021) Source: Statistics Austria Figure A2: Change in robot density 1993-2016 Source: International Federation of Robotics (IFR), own calculations 52 Table A1: Robotization and Employment - Additional Results By Sector By Occupation Manuf. Non-Manuf. Blue Collar White Collar (1) (2) (3) (4) Panel A: ∆ log(Employment) ×100 ∆ Robots -4.343 -0.031 -2.328 0.388 (1.683)** (1.178) (0.913)** (1.431) [0.299]*** [0.282] [0.14]*** [0.345] First-Stage F: 32.78 32.78 32.78 32.78 Panel B: %-change in Employment ×100 ∆ Robots -5.438 0.028 -2.134 -0.245 (1.585)*** (1.159) (0.946)** (1.358) [0.3]*** [0.265] [0.154]*** [0.314] First-Stage F: 32.78 32.78 32.78 32.78 Panel C: Separations (in % of initial employment) ×100 ∆ Robots -8.006 -15.938 (2.771)*** (4.712)*** [0.53]*** [1.322]*** First-Stage F: 32.78 32.78 Panel D: New Hirings (in % of initial employment) ×100 ∆ Robots -13.444 -18.072 (3.757)*** (5.054)*** [0.668]*** [1.373]*** First-Stage F: 32.78 32.78 Period Fixed Effects x x x x Region Fixed Effects x x x x Demographic Controls x x x x Regional Characteristics x x x x Labor Supply Shifts x x x x Labor Demand Shifts x x x x Detailed Industry Structure x x x x Commuting Zones 158 158 158 158 Periods 2 2 2 2 Observations 316 316 316 316 Notes: *<0.10, ** <0.05, *** <0.01. Conventional robust standard errors are shown in round brackets and shift-share robust standard errors from Adao, Koles´ar, and Morales (2019) are shown in square brackets. Units of observation are 158 clustered commuting zones (for details see Online Appendix C). All specifications include a set of region and period fixed effects, whereby the period fixed effects are interacted with the sum of exposure shares used to construct the explanatory variable (OLS) or the instrument (2SLS). Demographic controls include the start- of-period structure of the local workforce in 64 age-gender-education-nationality cells (demographic controls relating to the age-group 0-14 are not included in employment regressions). Regional characteristics control for the start- of-period logarithm of the gross regional product and the unemployment rate, as well as the start-of-period degree of urbanization. Shift-Share controls are included as the changes in import- and export-exposure and ICT-intensity (labor demand shifts) and changes in the migrant population differentiated by 4 educational groups (labor supply shifts). The detailed industry structure controls include start-of-period employment shares of several sub-industries of manufacturing (production of food products, consumer goods, industrial goods and capital goods), as well as industries outside of manufacturing (construction, personal services and business services) and the public sector. All regressions are weighted by start-of-period working-age population. 53 Online Appendix Online Appendix C Commuting Zones: This Section describes the construction of the commuting zones used as units of observation during the analysis. Thereby I strictly follow the methodology described in Tolbert and Sizer (1996) and Dorn (2009). The construction of commuting zones requires data on the commuting ties between municipalities. This data is taken from the register based Austrian census 2011 (available at Statistics Austria), and includes detailed information on municipality-to- municipality commuting flows. Denote municipalities with k= 1, ..., K. Then a K×Kcommuting-matrix is constructed, whereby rows iindicate the municipality of residence and columns jindicate the municipality of work. Each element of this commuting matrix therefore contains the number of workers who live in municipality iand work in municipality j, with the main diagonal indicating the number of workers who work in their residential municipality (i.e. do not commute). This commuting matrix is converted into a symmetric flow matrix. Following Tolbert and Sizer (1996), each element of this flow matrix is constructed as: Pij =Pji =fij +fji min(Li, Lj)(9) where fij denotes the absolute number of workers living in municipality iwho commute to municipality j, and Lidenotes the residential labor force of municipality i. Hence each element of the flow matrix Pcomputes as the sum of shared commuters between i and j, divided by the smaller residential labor force. As is explained in detail in Tolbert 60 and Sizer (1996), the symmetric flow matrix Pis characterized as a similarity matrix, where a higher value of Pij indicates a stronger commuting relationship between iand j. As the clustering algorithm (explained in detail below) requires a dissimilarity (or distance) matrix as input, the symmetric flow matrix is converted into a distance matrix D, whereby each element of this matrix computes as: Dij =Dji =       1−Pij if i6=j 0 if i=j (10) In this distance matrix, a lower value of Dij indicates stronger commuting ties (i.e. less distance) between municipalities iand j. In the main diagonal, where i=j, the distance is set to zero.28 To arrive at the desired commuting zones, a Hierarchical Cluster Algorithm is applied to the distance matrix D.29 This algorithm clusters elements of the distance matrix D, based on their average distance to each other, starting with the closest pair, and ending with one large cluster of all units. The algorithm stops clustering, once the average between cluster distance reaches a predefined threshold h. For the estimation of commuting zones for the US Tolbert and Sizer (1996) use an average between cluster distance of h= 0.98. Table A7 compares clustered Austrian commuting zones for different threshold values 28As Tolbert and Sizer (1996) mention in footnote 4 on page 12 of their paper, in some rare cases (i.e. whenever the sum of shared commuters between iand jis greater than the smaller one of the residential labor forces), Pij exceeds one. This would imply a negative distance between iand jin the distance matrix D. To avoid this, I follow Tolbert and Sizer (1996) and set Dij = 0.001 whenever this is the case. 29This type of algorithm is often used in machine learning applications and belongs to the broader class of unsupervised learning algorithms. 61 h, political districts and municipalities. As is to be expected, municipalities themselves perform rather poorly at containing commuter flows within their boarders with only about 47.3% of workers working in their residential municipality. Political districts, which are pre-defined administrative units, perform a little better. They contain around 65.6% of all commuters within the borders of 94 units (whereby the 23 districts of the capital Vienna are aggregated to a single observation). Table A7: Comparison of different local labor market (LLM) definitions LLM Commuters within LLM N Municipalities: 47.30% 2090 Political Districts: 65.62% 94 Commuting Zones: h = 0.98 70.07% 238 h = 0.9825 71.57% 197 h = 0.985 72.75% 158 h = 0.9875 74.18% 124 h = 0.99 75.31% 100 Note: Data on commuting flows is taken from the 2011 registry based census (available at Statistics Austria). To match municipality level data with other data sources used in the subsequent analysis, 2011 municipality structure has been mapped to the 2017 municipality structure. In cases were this mapping was ambiguous (i.e. when municipalities were split during reforms), some municipalities had to be aggregated to arrive at an municipality structure that is consistent over all used data sources. Therefore the number of municipalities slightly deviates from official sources. Comparing the performance of municipalities and political districts to clustered commuting zones, shows that the clustered commuting zones for all used thresholds h perform markedly better at containing commuting flows. For example the commuting zone for the clustering threshold h= 0.98 (i.e. the same threshold as in Tolbert and Sizer,1996) captures about 70% of commuting flows within its clusters. Compared to pre-defined political districts, it does this much more efficiently, as it captures a larger fraction of commuters (70% vs. 65.6%) in a higher number of clusters (238 vs. 94). Clustered commuting zones thus perform markedly better at controlling for spatial 62 shock spillovers, while also resulting in a higher number of available observations. This comes as no surprise, as the horizontal clustering algorithm is specifically designed to cluster regions according to their commuting ties. For the main part of the analysis, I rely on commuting zones clustered up to an average between cluster distance of h= 0.985. This threshold is slightly more restrictive than the one used by Tolbert and Sizer (1996) for the US in that it requires a greater average between cluster distance. Robustness of results to different clustering thresholds h To assess the influence of the tuning parameter hon the estimation results Tables A8 and A9 show corresponding estimations for several different values of this tuning parameter. Here Table A8 uses the baseline instrumental variable, and Table A9 uses the instrument were all Euro-Area countries are excluded from the computation. The very first configuration of the tuning constant (h= 0.98, column 2) corresponds to the configuration used for the clustering of US commuting zones in Tolbert and Sizer (1996) and Dorn (2009). All following configurations are more restrictive in that they allow weaker between-cluster commuting ties and thus result in a lower number of clusters. The baseline configuration which is used during the primary part of the analysis (h= 0.985) is presented in column 4. As is visible from Tables A8 and A9, the primary results of the analysis are robust to different configurations of the clustering algorithm. Here it stands out that more thoroughly controlling for spatial spillover effects (via higher configurations of h) mostly leads to an increase in effect size and precision. Since commuting zones which are clustered using higher values of the tuning constant hperform better at containing commuters within their borders, this indicates the presence of spatial spillover effects which potentially bias the estimates. This is further corroborated by Moran’s I test for 63 Table A8: Local Labor Market Definition (Baseline IV) (1) (2) (3) (4) (5) (6) Baseline LLM Definition: Districts h = 0.98 h = 0.9825 h = 0.985 h = 0.9875 h = 0.99 Commuters within LLM: 65.62 % 70.07 % 71.57 % 72.75 % 74.18 % 75.31 % Panel A: ∆ log(manufacturing employment) ∆ Robots 2.455 -5.407 -4.679 -4.343 -1.909 0.898 (1.082)** (0.902)*** (0.792)*** (1.683)** (1.421) (3.173) [0.616]*** [0.2]*** [0.176]*** [0.299]*** [0.32]*** [0.748] First-Stage F: 26.24 80.54 78.69 32.78 28.85 2.11 Moran’s I: 0.012 -0.032 0.002 0.031 -0.04 -0.019 (p-Value) (0.739) (0.14) (0.845) (0.174) (0.177) (0.683) Panel B: Net-outflow (all) ∆ Robots -0.159 0.051 0.388 1.055 0.242 0.172 (0.127) (0.164) (0.177)** (0.323)*** (0.314) (0.438) [0.025]*** [0.033] [0.035]*** [0.071]*** [0.077]*** [0.074]** First-Stage F: 13.22 78.31 80.9 29.77 17.73 2.77 Moran’s I: -0.109 0.023 -0.076 -0.026 0.008 0.033 (p-Value) (0.039)** (0.198) (0.002)*** (0.37) (0.686) (0.234) Panel C: Net-outflow (rural areas only) ∆ Robots 0.082 0.055 0.433 1.02 0.461 0.007 (0.102) (0.168) (0.188)** (0.342)*** (0.324) (0.505) [0.031]*** [0.034] [0.031]*** [0.069]*** [0.076]*** [0.094] First-Stage F: 13.22 78.31 80.9 29.77 17.73 2.77 Moran’s I: 0.074 0.01 -0.08 -0.028 0.002 0.024 (p-Value) (0.121) (0.529) (0.001)*** (0.333) (0.853) (0.38) Panel D: Net-outflow (rural to urban) ∆ Robots -0.07 0.023 0.23 0.565 0.483 -0.213 (0.054) (0.08) (0.087)*** (0.162)*** (0.14)*** (0.236) [0.009]*** [0.019] [0.013]*** [0.031]*** [0.03]*** [0.054]*** First-Stage F: 13.22 78.31 80.9 29.77 17.73 2.77 Moran’s I: 0.01 0.039 -0.058 -0.039 0.018 0.023 (p-Value) (0.748) (0.042)** (0.02)** (0.173) (0.45) (0.397) Full Controls x x x x x x Regions 94 238 197 158 124 100 Periods 2 2 2 2 2 2 Observations 188 476 394 316 248 200 Notes: *<0.10, ** <0.05, *** <0.01. Conventional robust standard errors are shown in round brackets, and shift-share robust standard errors from Adao, Koles´ar, and Morales (2019) are shown in square brackets. All specifications include a full set of control variables. All regressions are weighted by start-of-period working-age population. 64 Table A9: Local Labor Market Definition (Alternative IV) (1) (2) (3) (4) (5) (6) Baseline LLM Definition: Districts h = 0.98 h = 0.9825 h = 0.985 h = 0.9875 h = 0.99 Commuters within LLM: 65.62 % 70.07 % 71.57 % 72.75 % 74.18 % 75.31 % Panel A: ∆ log(manufacturing employment) ∆ Robots 1.788 -5.681 -4.622 -3.061 -2.618 -7.61 (1.065) (0.705)*** (0.719)*** (1.527)** (1.657) (2.268)*** [0.586]*** [0.185]*** [0.129]*** [0.224]*** [0.345]*** [0.55]*** First-Stage F: 50.25 158.56 179.06 58.02 42.02 4.16 Moran’s I: 0.026 -0.032 0.002 0.031 -0.043 -0.032 (p-Value) (0.535) (0.143) (0.853) (0.183) (0.145) (0.419) Panel B: Net-outflow (all) ∆ Robots -0.2 0.3 0.48 1.105 0.769 0.573 (0.111) (0.127)** (0.16)*** (0.257)*** (0.27)*** (0.291)* [0.023]*** [0.024]*** [0.032]*** [0.058]*** [0.054]*** [0.051]*** First-Stage F: 32.09 153.92 168.11 53.95 29.09 6.09 Moran’s I: -0.103 0.018 -0.076 -0.027 0.006 0.033 (p-Value) (0.051)* (0.293) (0.002)*** (0.366) (0.745) (0.231) Panel C: Net-outflow (rural areas only) ∆ Robots 0 0.286 0.484 0.934 0.728 0.286 (0.086) (0.141)** (0.172)*** (0.27)*** (0.282)** (0.334) [0.026] [0.025]*** [0.023]*** [0.056]*** [0.052]*** [0.072]*** First-Stage F: 32.09 153.92 168.11 53.95 29.09 6.09 Moran’s I: 0.054 0.006 -0.08 -0.028 0.003 0.023 (p-Value) (0.241) (0.667) (0.001)*** (0.341) (0.804) (0.385) Panel D: Net-outflow (rural to urban) ∆ Robots -0.067 0.114 0.223 0.435 0.433 -0.016 (0.047) (0.069) (0.079)*** (0.129)*** (0.125)*** (0.161) [0.008]*** [0.015]*** [0.01]*** [0.026]*** [0.022]*** [0.037] First-Stage F: 32.09 153.92 168.11 53.95 29.09 6.09 Moran’s I: 0.012 0.037 -0.058 -0.036 0.018 0.022 (p-Value) (0.721) (0.057)* (0.02)** (0.215) (0.454) (0.425) Full Controls x x x x x x Regions 94 238 197 158 124 100 Periods 2 2 2 2 2 2 Observations 188 476 394 316 248 200 Notes: *<0.10, ** <0.05, *** <0.01. Conventional robust standard errors are shown in round brackets, and shift-share robust standard errors from Adao, Koles´ar, and Morales (2019) are shown in square brackets. All specifications include a full set of control variables. All regressions are weighted by start-of-period working-age population. 65 spatial autocorrelation in the residuals which indicates the presence of spatial autocorrelation when using political districts (column 1) and less restrictive configurations of the clustering algorithm (columns 2 and 3). Importantly Moran’s I test fails to reject the null of no spatial autocorrelation for the configuration of the clustering algorithm that is used during the primary part of the analysis (h= 0.985, column 4), which suggests that all results in the main part of the paper are unaffected by spatial autocorrelation. Increasing the value of halso has an impact on the first stage regression, as the strength of the instrument declines with increasing h. Therefore the first stage F- statistic drops below 10 when increasing hto 0.99. Online Appendix D Shift-share standard errors: The shift-share robust standard errors proposed by Adao, Koles´ar, and Morales (2019) rely on asymptotic properties related to the number of industry shocks that are available to compute the shift-share variables. Correspondingly they may be downward biased, and thus overreject, in applications where the number of available industries is small. A possible downward bias due to an insufficient number of industry shocks is of particular concern in this application, because the IFR-data only contains information on 26 industries. This concern is further corroborated, since the AKM-standard errors are generally smaller than conventional heteroskedasticity robust standard errors. To address this concern, I apply an adjusted computation for the AKM-standard errors that uses the fact that each of the 26 IFR-industries is itself comprised of several sub-industries. Let any industry i= 1, ..., I be composed of several sub-industries jsuch that j= 1, ..., J ∈i. Then the shift-share variable for region rcan be re-written as: 66 ∆Robotsr=X i Empir Empr ∆Robotsi Empi (11) =X i X j∈i Empjr Empr!∆Robotsi Empi (12) because the employment share of industry iin region r is simply the sum of the employment shares of its sub-industries. To arrive at a shift-share expression for predicted regional exposure that uses the more detailed sub-industry based exposure shares one needs to assume that within industry iall sub-industries experience identical changes in robot density. Formally this assumption can be expressed as: ∆Robotsj Empj =∆Robotsi Empi ∀j∈i(13) In other words, this assumption states that every sub-industry jexperiences the same change in robots per worker, as is observed for the aggregated industry i. Using this assumption allows to re-write the shift-share expression in equation 11 as: ∆Robotsr=X iX j∈i Empjr Empr ∆Robotsj Empj (14) Notice that predicted robot exposure computed from expressions 11 and 14 result in the same value for regional robot exposure ∆Robotsr. The only difference between the two expression is that the first is calculated from industry-level exposure shares, while the second is calculated using more detailed sub-industry level exposure shares (and imposing the assumption laid out in equation 13). Therefore the same shift-share variable can be constructed from the exposure shares 67 of the aggregated industry itself: Empir Empr (15) or from the more detailed exposure shares of all sub-industries: Empjr Empr (16) whereby the second way results in a more detailed matrix of exposure shares which can be used in the estimation of the AKM-standard errors. Table A10 compares the AKM-standard error computation as proposed by Adao, Koles´ar, and Morales (2019) with standard error estimates for the described adjustment. Table A11 lists all available IFR-industries alongside the corresponding NACE Rev. 2 sub-industries used for the adjusted estimation. For illustrative purposes this comparison is carried out using estimations on the log-change in manufacturing employment. Other dependent variables result in a very similar picture (these results are available upon request). The comparison is performed for both variations of the AKM- standard error proposed in Adao, Koles´ar, and Morales (2019), whereby the first one (’AKM’) corresponds to the default30, and the second version (’AKM0’) corresponds to an estimate where the null is imposed in the estimation31. To account for the panel structure of the data the baseline estimates for both the AKM- and the AKM0-standard error cluster the available industries at the level of more aggregated industry categories (as is recommended for panel settings in Adao, Koles´ar, and Morales,2019). Because the adjustment described in equations 11 to 14 clearly violates the assumption of mutual independence of the robotization shocks (assumption 2 in Adao, Koles´ar, 30Equations 26 (OLS) and 36 (2SLS) in Adao, Koles´ar, and Morales (2019). 31Equation 27 (OLS) and comment below equation 36 (2SLS) in Adao, Koles´ar, and Morales (2019). 68 Table A10: Local Labor Market Definition Increase in SE Average Full Specification (1) (2) (3) (4) (5) (6) (7) (8) (9) OLS: ∆ Robots -0.551 -3.921 -3.711 -2.841 -4.137 -2.743 -0.885 AKM - W-matrix: standard (clustered) (0.045) (0.041) (0.077) (0.113) (0.107) (0.057) (0.032) AKM - W-matrix: sub-sectors (clustered) [0.361] [0.38] [0.518] [0.622] [0.521] [0.411] [0.474] + 706% + 1381% AKM - W-matrix: sub-sectors (unclustered) {0.333} {0.337} {0.45} {0.462} {0.476} {0.413} {0.438}+ 628% + 1269% AKM0 - W-matrix: standard (clustered) (0.045) (0.041) (0.077) (0.114) (0.107) (0.057) (0.032) AKM0 - W-matrix: sub-sectors (clustered) [0.362] [0.382] [0.521] [0.627] [0.526] [0.415] [0.478] + 646% + 1394% AKM0 - W-matrix: sub-sectors (unclustered) {0.333} {0.338} {0.45} {0.463} {0.477} {0.414} {0.439}+ 628% + 1272% 2SLS: ∆ Robots -1.185 -5.984 -6.254 -3.823 -4.006 -3.599 -4.343 AKM - W-matrix: standard (clustered) (0.241) (0.301) (0.312) (0.252) (0.21) (0.137) (0.095) AKM - W-matrix: sub-sectors (clustered) [4.337] [0.952] [0.809] [0.554] [0.476] [0.328] [0.237] + 373% + 149% AKM - W-matrix: sub-sectors (unclustered) {3.862} {0.713} {0.626} {0.509} {0.495} {0.394} {0.299}+ 340% + 215% AKM0 - W-matrix: standard (clustered) (0.243) (0.301) (0.313) (0.252) (0.21) (0.137) (0.095) AKM0 - W-matrix: sub-sectors (clustered) [4.337] [0.952] [0.809] [0.554] [0.476] [0.328] [0.237] + 371% + 149% AKM0 - W-matrix: sub-sectors (unclustered) {4.048} {0.717} {0.629} {0.512} {0.498} {0.395} {0.3}+ 350% + 216% Period Fixed Effects x x x x x x x Region Fixed Effects x x x x x x x Demographic Controls x x x x x x Regional Characteristics x x x x x Labor Demand Shifts x x x x Labor Supply Shifts x x x Manufacturing Share x Detailed Industry Structure x Commuting Zones 158 158 158 158 158 158 158 Periods 2 2 2 2 2 2 2 Observations 316 316 316 316 316 316 316 Notes: Significance stars omitted for readability. Units of observation are 158 clustered commuting zones (for details see Online Appendix C). All specifications include a set of region and period fixed effects, whereby the period fixed effects are interacted with the sum of exposure shares used to construct the explanatory variable (OLS) or the instrument (2SLS). Demographic controls include the start-of-period structure of the local workforce in 64 age-gender-education-nationality cells (demographic controls relating to the age-group 0-14 are not included in employment regressions). Regional characteristics control for the start-of-period logarithm of the gross regional product and the unemployment rate, as well as the start-of-period degree of urbanization. Shift-Share controls are included as the changes in import- and export-exposure and ICT-intensity (labor demand shifts) and changes in the migrant population differentiated by 4 educational groups (labor supply shifts). The detailed industry structure controls include start-of-period employment shares of several sub-industries of manufacturing (production of food products, consumer goods, industrial goods and capital goods), as well as industries outside of manufacturing (construction, personal services and business services) and the public sector. All regressions are weighted by start-of-period working-age population. 69