Robots and Jobs Revisited: A Bit-Exact Replication, Shift-Share Anatomy, and Capital-Stock Extension of Acemoglu and Restrepo (2020)

Abstract

Version 1.01, 16 September 2026. This paper revisits the robot-exposure design of Acemoglu and Restrepo (2020). The first contribution is a bit-exact replication of a deliberately narrow benchmark: Table 7, columns 3 and 6. Starting from the underlying data sources, I reconstruct the inputs needed for these two baseline long-difference IV estimates: commuting-zone employment outcomes, commuting-zone-by-demographic-group wage outcomes, baseline controls, robot exposure measures, and instrumental-variable specifications. The replicated coefficients match the published benchmark. One additional robot per thousand workers is associated with a decline of 0.388 percentage points in the private employment-to-population ratio and a 0.768 percent decline in log hourly wages. The second contribution is diagnostic. I ask whether the replicated estimate is fragile in the ordinary empirical sense. It is not. Standard robustness checks—alternative controls, weights, outlier treatments, and sample perturbations—leave the main qualitative result broadly intact. This matters because it separates two questions that are often conflated. A result can be stable across conventional specifications and still rely on a narrow source of identifying variation. The third contribution examines that identifying variation directly. The Acemoglu–Restrepo design is a shift-share IV design: local exposure is built from industry-level robot-adoption shocks interacted with pre-existing local industry shares. Rotemberg-weight, leave-one-industry-out, and commuting-zone identifying-mass diagnostics show that the baseline estimate is strongly concentrated in automotive exposure. This concentration is not specific to the baseline five-country European average. Reconstructing Europe-5 plus Germany and the full nine-country European set leaves both employment and wage estimates negative and produces much stronger regional first stages, while automotive’s normalized absolute Rotemberg mass rises from about 84 percent to 89 and 92 percent, respectively. The non-automotive instrument does not sustain the original negative estimate with a strong first stage. Using the authors’ underlying SIC87DD trade data, I also construct automotive-specific exposure to Mexican trade and to trade from Germany, Japan, and South Korea. These finer trade controls do not attenuate the negative robot estimates. Instead, the estimates become more negative while first-stage strength falls, showing that automotive robot and trade shocks draw on closely overlapping local industrial support. A complementary timing-window frontier shows that this concentration should not be confused with fragility to the particular dates chosen by the original study. I reconstruct all 78 admissible European industry-shock windows between 1993 and 2007 with at least three elapsed years, while holding the outcomes, endogenous U.S. exposure measure, controls, weights, fixed effects, clustering convention, and exact estimation samples fixed. The baseline 1993–2007 window is not an outlier. Among the 60 windows with a Kleibergen–Paap F-statistic of at least 10, all produce negative and statistically significant employment and wage estimates. Their median coefficients, −0.370 for employment and −0.740 for wages, are close to the baseline estimates of −0.388 and −0.768. Yet these strong alternative instruments remain highly concentrated in automotive exposure. The timing choice is therefore not the main empirical vulnerability. The more important issue is the exclusion restriction and the external interpretation of an automotive-led shift-share estimand. A fourth contribution is inferential. I rewrite the baseline shift-share IV as an algebraically equivalent IV at the level of the underlying industry shocks and independently reproduce the corresponding BHJ shock-level implementation. The headline coefficients and robust shock-level standard errors match the original replication code. Translating the regional first stage to the shock level yields BHJ exposure-robust SSIV first-stage F-statistics of 6.23 for employment and 6.33 for wages, substantially below the regional Kleibergen–Paap statistics of 33.62 and 34.35. With only 19 industry shocks, the finite-sample treatment of inference matters. Asymptotic heteroskedasticity-robust Anderson–Rubin tests yield p-values of 0.033 and 0.035, while the small-sample F versions yield p-values of 0.059 and 0.063. Thus the asymptotic tests reject a zero effect at the five-percent level, whereas the small-sample versions do not. Treating services, whose Europe-5 shock is exactly zero, as outside the active shock set raises the corresponding BHJ first-stage F-statistics to 9.52 and 9.92, but the small-sample Anderson–Rubin p-values remain 0.101 and 0.068. Predetermined U.S. industry trajectories also show evidence of association with subsequent Europe-5 robot shocks. Conditioning on pre-1990 U.S. industry employment growth leaves the coefficients negative but produces small-sample Anderson–Rubin p-values around 0.08–0.10. The fifth contribution is constructive and diagnostic. I reconstruct robot exposure using an age-adjusted perpetual-inventory stock anchored to the 1993 IFR operational stock and subsequently updated using installation flows under explicit depreciation assumptions. The baseline Acemoglu–Restrepo measure already uses changes in reported operational robot stocks; the extension therefore does not replace flows with stocks. Instead, it makes the persistence and depreciation structure of robot vintages explicit. The negative employment result survives this alternative measurement. However, the capital-stock treatment does not broaden the identifying support of the instrument. Across the raw-stock and PIM specifications, automotive accounts for approximately 86–90 percent of normalized absolute Rotemberg weight, while the three largest industries account for 94–97 percent. Removing the automotive component reduces the Kleibergen–Paap first-stage statistic to roughly 1.6–2.1 across specifications. Thus, alternative vintage weighting changes the measurement of robot exposure but does not resolve the concentration of identifying variation in automotive. The paper’s main conclusion is simple. The original estimates are computationally reproducible and conventionally robust. But their interpretation depends on where the shift-share design obtains its identifying support. Conventional robustness shows that the estimate survives ordinary specification changes. Shift-share anatomy shows what kind of variation makes it survive. JEL Codes: J23, J24, O33, R23, C26 Keywords: Industrial robots; automation; employment; wages; replication; computational reproducibility; shift-share IV; Bartik instruments; Rotemberg weights; local labor markets; commuting zones; automotive industry; technological change; robot capital stock; perpetual inventory method; weak identification; shock-level inference