Part 1 makes four major contributions. First, we develop a new measure of automation in machinery. We classify technology codes (CPC/IPC codes) using the text of some patents. As IPC codes are universal, we can then classify all patents. This classification is publicly available, and several researchers have used it for their own projects. Second, we map patents to their industry of use, using data on the inventing industry and a capital flow table (an input/output table for capital goods). We find that industries exposed to a higher share of automation patents in machinery experience a decline in routine tasks, a decline in the labor share, and an increase in the ratio of high- to low-skill workers. Third, we use global firm-level data to analyze the causal effect of an increase in labor costs on automation innovations. We measure innovators’ exposure to each country using patent data and compute the low- and high-skill labor costs faced by their customers as a firm-specific weighted average of country-level labor costs. We control for (main) country-year fixed effects, ensuring that we compare how firms from the same country react to wage-induced foreign demand shocks. We find a positive effect of low-skill wages on automation innovation with an elasticity of 2-5 and a negative effect of high-skill wages. There is no effect on non-automation innovation performed by the same set of firms. An increase in minimum wages also induces automation innovation. Fourth, we look at the German Hartz reforms, which reduced low-skill labor costs. We show in an event-study that these reforms reduced automation innovations of non-German firms highly exposed to Germany. We presented this paper in many venues, and it is now published at the Journal of Political Economy.
Part 2 leverages the classification of Part 1. We use more detailed disaggregated US sectoral data from 1980 to 2010 and compute decadal change in employment. We find that overall machinery innovations used to increase employment but now decrease it. Decomposing these into automation and non-automation innovations, the former decrease employment while the latter increase it. An industry which is 1 standard deviation more exposed than another one to automation experiences a decrease in employment of 16% per decade, which is equal to the average employment decrease in manufacturing per decade. We then move to the commuting zone (CZ) level. We measure local automation exposure using a shift-share strategy. The CZ results mirror the industry level results but allow us to go further: we show that the non-manufacturing sector does not compensate for manufacturing employment losses from automation, and that routine occupations are particularly affected. Finally, we show that automation innovations do not generate compensating employment gains for the producing industries or the CZ where innovations take place. We have derived the empirical results and are about to write the draft.
In Part 3, we find that the decline in the aggregate labor share is mostly a within industry phenomenon: a larger share of the industry value added is produced by low labor share firms, while the labor share of the median firm increases. This reallocation is driven by firms that at the same time become low labor share and increase their size. While we do not rule out a role for automation, our analysis emphasizes the role of trade: the decline in the labor share is more pronounced in export-oriented sectors, and low labor share firms grow by expanding their sales abroad. R&D investment predicts declines in the labor share, suggesting that a plausible channel is that successful innovators, who can charge a large mark-up, are now able to capture a larger market.
This part is still ongoing: We have established the stylized facts, we did some preliminary work on the model, but we are not done yet.
