Robots at work? : pitfalls of industry level data

In a seminal paper Graetz and Michaels (2018) find that robots increase labor productivity and TFP, lower output prices and adversely affect the employment share of lowskilled labor. We demonstrate that these effects are heavily influenced by the sample composition and argue that focusing on manufacturing and mining sectors mitigates unobserved heterogeneity and is more coherent with an identification strategy that rests on instruments that do not vary by industries. In sum, this leads to more plausible results regarding the overall economic effects of robotization, whereby the focus on robotizing industries leads to a sizable drop of the productivity effects, halving the effect size for labor productivity and insignificant price effects. The most pronounced consequences from the sample choice occur for labor market outcomes, where significant negative employment effects become insignificant and positive wage effects are reversed into the opposite. We show that controlling for demographic workforce characteristics is essential for obtaining significant labor productivity effects and leads to the negative effects of robots on wages. Additionally, investigating only robotizing sectors does not corroborate skill-biased technological change due to robotization, but rather, indicates towards labor market polarization. Finally, we document a non-monotonicity in one of the instruments, which calls for caution in the use of that instrument.