search

chaisemartinPackages / did_multiplegt

Estimation in Difference-in-Difference (DID) designs with multiple groups and periods.
7 stars 3 forks source link

Different sign for placebo coefficient with dyn and stat (Stata) #26

Closed Gilles-Spielvogel closed 1 week ago

Gilles-Spielvogel commented 2 months ago

Hello,

Many thanks for the library and the recent updates!

I have estimated (almost) the same model on the same data (continuous outcome, binary treatment, two periods) using the dyn and stat modes. I find as expected (almost) the same results for the DID estimator, but not for the placebo regression. In fact, the absolute value of the placebo coefficient is the same, but the sign is reversed. I find this puzzling and I thought this was an issue worth raising with you.

My models and results look like this:

did_multiplegt (stat) Y G T D, estimator( as) cluster(G) switchers(up) exact_match placebo(1) graph_off

Estimate SE LB CI UB CI Switchers Stayers
AS 4.213364 .8534473 2.540607 5.88612 500 11786
Placebo_1 -.892202 .7020076 -2.268137 .4837329 474 5656

did_multiplegt (dyn) Y G T D, cluster(G) switchers(in) only_never_switchers placebo(1) graph_off

Estimate SE LB CI UB CI N Switchers
Effect_1 4.249127 .7965185 2.687979 5.810274 11811 499
Placebo_1 .8922021 .7030345 -.4857202 2.270124 6130 474

Many thanks if you can help me understand this!

Best,

Gilles

chaisemartinPackages commented 1 month ago

Dear Gilles,

Thank you Gilles for this test, this is all the more reassuring !

This derives from different conventions across packages. In de Chaisemartin et D’Haultfœuille 2024, placebos are constructed by taking differences that go in the past (see section 3.5) compared to the first treatment period. On the contrary in de Chaisemartin et al. 2024, placebos are forward differences before the period of the change.

This explains why the coefficients can be similar but of opposite signs.

Hope this helps, Romain