I believe the paper discussed is here [0]. I've skimmed it and am somewhat underwhelmed by both the methods and theory: * All of their regressions are in level

I believe the paper discussed is here [0]. I've skimmed it and am somewhat underwhelmed by both the methods and theory: * All of their regressions are in levels. * Look at Table 1, Column 1. * Robotics investment is positively correlated with overall employment. * What this tells me is that, "Large companies, by dint of their being large, make large investments. Some of those investments are in robotics. Likewise for small companies." * To me, it has nothing to do with the claim "More robots => more employment." * Running regressions in first differences would be better (but not enough). * Then you could begin to work towards an interpretation like, "companies that _increased_ their investment in robotics in year X tended to increase/decrease employment/management/turnover by X." You can account for lags and the like via your specification. * Fixed effects _are not_ equivelant to first differences in this setup! * Separate out the long-, medium- and short-term consequences: * Investment in new technology may require hiring new "talent" that is knowledgeable of the tech. * In cases where there is a tech shock and firms are all buying the same machine, demand for a certain skillset will grow quickly. So you'd likely see: * An uptick in turnover among affected workers * A move to a "flatter" organizational structure (various reasons). So maybe fewer people who are managers per se. * Possibly short-term productvity decreases arising from implementation issues. (e.g. have to hire people to check the work of machine in staging manually while still running previous process.) * Over time, though: * Implementation issues/bugs are smoothed out * Rare skills are commidified. The need to hire "superstars" ends. * It is important to ask, "what might have caused investment by a given firm over a given time frame?" * Common industry technological shocks might spur investment. * (And the simplest way to account for technological shocks in this sort of regression you'd want `[industry] x [year]` fixed effects.) * "We went on a manager hiring spree in year 1. Those managers, in turn, bought a lot of robots in year 2. In year 2 we hired fewer managers because we'd just hired a bunch in year 1." * Reverse causation issues are pervasive. * Finally, I think it is important to notice their notion of "robot adoption": * They are talking about physical hardware _only_. * Their measure of investment is coming only from import data. The set of firms that are going to go through the rigammarole of importing physical hardware is probably raher special: Probably: * Larger * Have a high expected ROI from investment At the very least, this paper needs first-difference regressions and a good instrument that can help to disentangle what might be going on. For the instrument, it might make sense to look into changes in tariff duties over time [1]. --- [0] https://content.tcmediasaffaires.com/LAF/lacom2019/robots.pdf https://content.tcmediasaffaires.com/LAF/lacom2019/robots.pd... [1] https://cbsa-asfc.gc.ca/trade-commerce/tariff-tarif/2020/html/tblmod-2-eng.html https://cbsa-asfc.gc.ca/trade-commerce/tariff-tarif/2020/htm...