Key points: the estimator is
- robust to heterogeneous effects
- efficient under homoskedasticity
- transparent
- robust to pre-testing
- works w/controls, triple-diffs, etc
With event_plot, you can make event study plots for at least 5 methods:
- our did_imputation
- robust estimators of @CdeChaisemartin-D’Haultfoeuille, Callaway-@pedrohcgs, Sun-Abraham
- traditional OLS
And combine them with each other if you wish!
On this figure you can see:
- how precise each event study estimator is (which of course depends on the DGP)
- how long estimation takes
- how dynamic OLS is biased with heterogeneous effects (Sun-Abraham’s result)
This is likely the first time all estimators are on one graph. The code is on github for you to use in any application!
But don’t overlook the differences in assumptions, properties & interpretation of coefs (e.g. which reference groups are used for causal effects & pre-trends)
Comments and feedback on both commands and on our paper with @XJaravel and @jannspiess are most welcome!
And special thanks to @kylefbutts for the help with preparing the helpfiles for both commands
• • •
Missing some Tweet in this thread? You can try to
force a refresh
📢 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
@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
What do we do?
1) Clarify the assumptions behind event studies 2) Discuss problems with conventional OLS estimation, arising because the assumptions get conflated 3) Derive a robust & efficient "imputation" estimator, inference for it & a new pre-trend test (Stata code included)
- 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!)
We discuss a likely bias in such settings due to non-random exposure to exogenous shocks
Imagine randomly connecting regions by roads, as in an RCT. These random shocks increase market access. But a regression on MA growth may suffer from omitted varables bias (OVB). Why?