Thread by @glad_1723

Are you using the Shift-Share Design?

Today, we read the new @QJEHarvard by Adao, Kolesar & Morales who show us that many Shift-Share Designs produce too small confidence intervals (i.e. many false positives). A must-read for applied economists!

#econfriday

In the Shift-Share Design, we identify a causal effect via two components: An exogenous shock to all observations in our sample (the "Shift") combined with how much an observation is hit by that shock (the "Share").

To compute the "Share", we usually rely on pre-determined characteristics of a unit. For a subnational analysis of the effects of a global price-shock to some good for example, we might compute how much a region produced of this good before the shock.

The problem: Regions with similar Shares (e.g. sectoral compositions) are likely affected by the same (sector-level) unobserved shocks. Consequently, residuals of regions with similar shares will likely be correlated.

Does it matter? The authors use employment-shares by sector for US municipalities and estimate a Shift-Share regressor using random shocks to randomly chosen sectors. Even though such random shocks should have no effect at all, they find 5%-significance in up to 55% of the cases!

The rescue: To leave us on a positive note, the authors provide a solution to the problem. They present a new framework which allows for unobserved sectoral correlation in the residual.

-> R package: github.com/kolesarm/Shift…

Concluding: A great paper with an incredibly important advice to all applied economists (curious about) using Shift-Share designs! Thoughtful explanations and helpful illustrations of the problem make it a well-written and super helpful publication!

economics.harvard.edu/files/economic…

..and there is much more interesting research on shift-share design!

Borusyak, @autoregress & Jaravel recommend running/vizualising regressions at the shock level (where the identifying variation comes from):
mit.edu/~hull/bartik_0…

-> Stata package: mit.edu/~hull/bartik_0…

@paulgp, Sorkin & Swift explore the instances when the shares are assumed to be exogenous and show how to explore how much weight a specific instrument has
paulgp.github.io/papers/bartik_…

-> Stata package: github.com/paulgp/bartik-…

These papers are all great and expand our knowledge on shift-share designs! Thank you all - especially for providing all these great packages as well. We might cover the latter two in the future in more detail.

To complete the thread: 2 more very interesting papers that relate to shift-share IVs

@DavidAJaeger, Ruist & Stuhler caution that these IVs can conflate long & short run effects if these shocks are serially correlated and affect the same units over time

nber.org/papers/w24285.…

Christian & @cbb2cornell stress that researchers need to check for long run trends in the time series component of the IV and the outcome since these can bias the effect

bit.ly/31ehFaG

Are there more recent papers that relate to the debate?