🚨 Difference-in-differences working paper alert 🚨
Our Local-Projections DiD offers a unified approach that encompasses many popular alternatives as specific instances; allows for extensions; and does it all using an OLS regression.
nber.org/papers/w31184
🧵
Our Local-Projections DiD offers a unified approach that encompasses many popular alternatives as specific instances; allows for extensions; and does it all using an OLS regression.
nber.org/papers/w31184
🧵
Why yet another DiD estimator?
* Simple, fast, easy to code (single regression command with an "if" statement)
* Flexible (eg, recurring treatments)
* Easily allows matching on pre-treatment outcomes and covariates
2/
* Simple, fast, easy to code (single regression command with an "if" statement)
* Flexible (eg, recurring treatments)
* Easily allows matching on pre-treatment outcomes and covariates
2/
Baseline version of LP-DiD: estimate a LP regression limiting to "clean controls"
3/
3/
Baseline, unweighted, approach identifies a VW ATT.
Identical to the "stacked approach" of Cengiz Dube Lindner Zipperer ('19). But can do it without stacking!
4/
Identical to the "stacked approach" of Cengiz Dube Lindner Zipperer ('19). But can do it without stacking!
4/
But that's just the beginning...LP-DiD can do many things!
Example: we derive the VWATT weights, so can easily re-weight & estimate an equal weighted ATT. We show the re-weighetd LP-DiD is numerically identical to Callaway and Sant'Anna ('20) but very easy/fast to compute.
5/
Example: we derive the VWATT weights, so can easily re-weight & estimate an equal weighted ATT. We show the re-weighetd LP-DiD is numerically identical to Callaway and Sant'Anna ('20) but very easy/fast to compute.
5/
And if you reweight + use a broader pre-treatment period for differencing the outcome, the LP-DiD estimate is shown to be very close to the Borusyak, Jaravel, Spiess imputation estimator.
6/
6/
You want more? We got more.
LP-DiD easily takes covariates (of course it does, it's an OLS regression!), including pre-treatment outcomes. The latter can be useful for constructing a counterfactual that guards against certain types of violations of parallel trends.
7/
LP-DiD easily takes covariates (of course it does, it's an OLS regression!), including pre-treatment outcomes. The latter can be useful for constructing a counterfactual that guards against certain types of violations of parallel trends.
7/
Give us even more, you say? OK, you got it!
Easily incorporates continuous treatment (with additional assumptions needed for interpretation) or non-absorbing treatments (like minimum wages) where truly never-treated units rarely exist.
8/
Easily incorporates continuous treatment (with additional assumptions needed for interpretation) or non-absorbing treatments (like minimum wages) where truly never-treated units rarely exist.
8/
For more, including performance in terms of MSE, bias, as well as computational speed, check out the paper!
Joint with @DanieleGirardi_, O Jorda, and A Taylor.
Also see Daniele's Slide Deck: danielegirardi.github.io/posts/document…
9/
Joint with @DanieleGirardi_, O Jorda, and A Taylor.
Also see Daniele's Slide Deck: danielegirardi.github.io/posts/document…
9/
Finally, for STATA example code on implementing the LP-DiD, check out the Github repository here:
github.com/danielegirardi…
10/
github.com/danielegirardi…
10/
11/ Tagging a few individuals who may be interested: @pedrohcgs @borusyak @XJaravel @jmwooldridge @CdeChaisemartin @agoodmanbacon @instrumenthull @paulgp
• • •
Missing some Tweet in this thread? You can try to
force a refresh
