1 The Causal Question
Does the adoption of industrial robots reduce employment and wages? The question sounds simple but is empirically treacherous. Firms that adopt robots are not randomly selected. Industries that automate tend to be those facing rising wages—creating a reverse-causality problem that naive OLS would not solve. Aggregate time-series regressions cannot separate the effect of robots from the many other forces reshaping labour markets over the same period.
Acemoglu and Restrepo [2020] addressed this challenge with an identification strategy borrowed from Autor et al. [2013] and used it to estimate the causal effect of industrial robot adoption on employment and wages across US commuting zones. Published in the Journal of Political Economy, the paper became one of the most cited contributions on the economics of automation.
2 The Theoretical Framework
Acemoglu and Restrepo situate the empirical work within a model of task-based production. In the standard model, workers and capital perform tasks that combine into output. When a new technology such as a robot automates a task previously performed by labour, it displaces workers from that task. Whether total employment rises or falls depends on whether the productivity gains create enough new tasks (the reinstatement effect) to offset the displacement.
This framework predicts that robot adoption reduces employment and wages in the short run through task displacement, but that in the long run, complementary tasks may be created. The empirical question is about the medium-run: over the period 1990-2007, was the net effect positive or negative?
3 The Identification Strategy
3.1 The Shift-Share Instrument
The core identification challenge is that robot adoption is endogenous. US regions that adopted robots quickly may have done so because their industries were expanding, or because local wage costs were rising. Either story generates a spurious relationship between robots and employment that does not reflect the causal effect of automation.
The paper constructs a Bartik-style shift-share instrument. For each commuting zone c, exposure to robots is proxied by:
where Lck,1990 / Lc,1990 is the employment share of industry k in commuting zone c in 1990, and ΔRobotsk,Europe is the change in robots per thousand workers in industry k in six European countries (Germany, France, Italy, Sweden, Denmark, Finland) over the same period.
The intuition is straightforward. A US commuting zone with a high share of employment in auto manufacturing (which adopted many robots in Europe) is predicted to be exposed to more robots than one specialised in food services (which did not). But the instrument uses European adoption rates rather than US adoption rates, isolating the component of US robot exposure driven by global technological trends in each industry rather than by local US labour market conditions.
3.2 Validity of the Instrument
The shift-share instrument's validity rests on two distinct arguments, corresponding to the two components of the instrument:
- Share exogeneity: The 1990 industry composition of US commuting zones must be unrelated to subsequent labour market trends, conditional on a rich set of controls. Acemoglu and Restrepo include controls for the initial share of routine jobs, the initial share of employment in manufacturing, exposure to import competition from China [Autor et al., 2013], and demographic variables.
- Shock exogeneity: The European robot adoption rates must reflect global technological trends rather than anything driven by US conditions. This is plausible because European labour markets face different institutional environments, union structures, and minimum wage levels than the US—factors that would not systematically induce European firms to adopt robots in industries where US workers are struggling.
The two frameworks for interpreting shift-share validity—Goldsmith-Pinkham et al. [2020] who emphasize shares and Borusyak et al. [2022] who emphasize shocks—both find support in this application.
3.3 Data Sources
Robot data come from the International Federation of Robotics (IFR), which reports annual counts of industrial robots by industry and country. The IFR defines industrial robots as "automatically controlled, reprogrammable multipurpose manipulators"—the assembly-line robots used in welding, painting, material handling, and packaging. The IFR data are matched to US industry classifications. Labour market outcomes—employment-to-population ratios, wages, industry composition—come from the Census and the Current Population Survey at the commuting zone level. The sample spans 722 US commuting zones and two periods: 1990-2000 and 2000-2007.
4 Key Findings
4.1 First Stage
European robot adoption is a strong predictor of US commuting zone robot exposure, conditional on controls. The first-stage F-statistic comfortably exceeds conventional weakinstrument thresholds. One additional robot per thousand workers in European industriestranslates into significantly higher robot adoption in US commuting zones with heavy employment shares in those same industries.
4.2 Reduced-Form and IV Estimates
The main finding is strikingly negative. One additional robot per thousand workers reducesthe employment-to-population ratio by 0.20 percentage points and reduces wages by 0.42percent [Acemoglu and Restrepo, 2020]. These are meaningful magnitudes. The US adopted3roughly 0.35 additional robots per thousand workers over 1990–2007, implying that robotadoption reduced the employment-to-population ratio by about 0.07 percentage points andwages by about 0.15 percent—modest in aggregate but concentrated in the commuting zonesmost exposed to automation.
The disaggregation reveals that employment losses are concentrated in manufacturingand among workers without college degrees. Commuting zones in the Midwest and Southeast, with high initial employment shares in auto manufacturing and electronics assembly,experienced the largest employment declines.
4.3 Mechanisms
The paper distinguishes task displacement from productivity effects. Using industry-leveldata, Acemoglu and Restrepo show that robot adoption reduces employment within theindustries that adopt them (within-industry displacement) and has spillovers to other industries in the same local economy. The productivity channel—through which higher outputgenerates demand for other local goods and services—appears insufficient to offset the displacement effect in the medium run.
5 Limitations
5.1 General Equilibrium
The commuting-zone analysis estimates a partial equilibrium effect: the impact on a locallabour market treated as a small open economy. In general equilibrium, higher productivityfrom robots in some sectors may raise wages nationally. A full accounting of welfare effectswould require a structural trade model, as in Caliendo et al. [2019], to trace the economy-wideeffects of automation through input-output linkages and price adjustments.
5.2 Instrument Scope
Industrial robots are a narrow technology—about 0.4 million units installed in the US by2007, compared to the tens of millions of software applications, algorithms, and AI systems that now affect labour markets. The findings are identified off variation in the specificindustries that adopted IFR-defined robots, which are disproportionately in heavy manufacturing. The external validity to broader automation—including software, AI, and logisticsautomation—is limited.
5.3 Dynamic Effects
The medium-run estimates (1990–2007) may not capture the full long-run adjustment. Ifworkers displaced by robots retrain, migrate, or shift to complementary tasks, the long-runemployment effect could be smaller than the medium-run estimates suggest. The reinstatement channel, which is central to the theoretical framework, operates over decades ratherthan years.
6 What We Learn
The Acemoglu-Restrepo paper establishes three lessons. First, industrial robot adoption hasa negative causal effect on employment and wages in exposed commuting zones. Second,the magnitude is meaningful but not catastrophic in aggregate—robot adoption explains amodest fraction of total labour market changes over 1990–2007. Third, the effects are deeplyunequal: concentrated in specific industries, skill groups, and geographic areas, contributingto the widening spatial inequality that characterises American labour markets today [Autor,2019].
Methodologically, the paper extends the Autor-Dorn-Hanson shift-share approach to anew technology shock, demonstrating the versatility of this identification strategy while alsoillustrating its limitations—the instrument’s validity depends on the exogeneity of Europeanadoption rates, which can be contested, and on the correct specification of the first-stageindustry shares.
References
- Acemoglu, D. and Restrepo, P. (2020). Robots and jobs: Evidence from US labor markets. Journal of Political Economy, 128(6):2188-2244.
- Autor, D. H., Dorn, D., and Hanson, G. H. (2013). The China syndrome: Local labor market effects of import competition in the United States. American Economic Review, 103(6):2121-2168.
- Autor, D. H. (2019). Work of the past, work of the future. AEA Papers and Proceedings, 109:1-32.
- Borusyak, K., Hull, P., and Jaravel, X. (2022). Quasi-experimental shift-share research designs. Review of Economic Studies, 89(1):181-213.
- Caliendo, L., Dvorkin, M., and Parro, F. (2019). Trade and labor market dynamics: General equilibrium analysis of the China trade shock. Econometrica, 87(3):741-835.
- Goldsmith-Pinkham, P., Sorkin, I., and Swift, H. (2020). Bartik instruments: What, when, why, and how. American Economic Review, 110(8):2586-2624.