🧵 Do shift-share IV regressions suffer from negative weight problems?

@CdeChaisemartin and Lei recently posted a WP (Worrying Paper, one could say) arguing that way

@instrumenthull and I are more optimistic and decided to share our view in a brief note
https://t.co/J6dUwHyMWkdropbox.com/scl/fi/vi0jkwo…
We make 4 points about shift-share IVs under heterogeneous effects:

1/4) They identify a convexly weighted average of causal effects when the shocks are as-good-as-random

It's like a LATE result, and our ReStud paper w/@instrumenthull @XJaravel proves it in more generality

Here's my version of this very important point about negative weights:

You should consider the nature of treatment assignment and reasonable restrictions on treatment effects before worrying about negative weights, especially when you can’t avoid them by imputation

➡️🧵

https://twitter.com/instrumenthull/status/1682084118333358080

There are two types of weights in regressions: "ex post" (conditionally on treatment realizations) and "ex ante" (before treatments are realized)

Ex post weights are negative for some treated observations for almost every regression: some obs in the comparison group are treated

📢 I’d like to share with #TradeTwitter a 🧵 on what @XJaravel and I have learned about the unequal effects of international trade through both cost-of-living and wages in the U.S.

For those of you who have seen my JMP, this is a much-revised draft

dropbox.com/s/eiygfth61vp4… @XJaravel Let’s start with the effects on costs-of-living, which are understood less well

Who benefits more from lower prices of imports in the US?

A typical guess & prior estimates: poor consumers who buy more tradable goods, esp. from China

So trade could REDUCE (real) inequality

On the 1st anniversary of my twitter account, I'm happy to share w/the #EconTwitter community our #stata commands for event studies:

did_imputation: robust and efficient imputation estimator

event_plot: event study plots after various estimation methods

github.com/borusyak/did_i… I’ve discussed did_imputation in the 🧵 on our paper w/@XJaravel and @jannspiess

Key points: the estimator is
- robust to heterogeneous effects
- efficient under homoskedasticity
- transparent
- robust to pre-testing
- works w/controls, triple-diffs, etc

https://twitter.com/borusyak/status/1394680063384248320

🥁 I’m thrilled to announce our paper w/@XJaravel and @jannspiess, “Revisiting Event Studies: Robust and Efficient Estimation” 🥁

It’s a fully revised version of our 2017 draft that the diff-in-diff loving audience may have seen

dropbox.com/s/y92mmyndlbku… @XJaravel @jannspiess Event studies are diff-in-diff research designs with staggered adoption of treatment:

- Units in a panel get treated at different times (and stay treated forever)
- Some units may be never treated

We are interested in heterogenous or dynamic treatment effects

New WP: “Non-Random Exposure to Exogenous Shocks” (w/ @autoregress). Summary 🧵:

Papers often estimate causal effects by leveraging exogenous shocks that affect many observations jointly, to different extents

We show problems w/this & offer new solutions
dropbox.com/s/brhuxe1b1k8x… Examples:

- New railroad construction affects all regions via increased market access

- State-level policies affect different people depending on demographics

- Trade shocks affect different regions

- Shocks propagate through networks.

(More in the paper!)