search

xuyiqing / gsynth

Generalized Synthetic Control Method
Other
132 stars 40 forks source link

Individual unit effect inference for matrix completion #28

Open elandesberg opened 4 years ago

elandesberg commented 4 years ago

@xuyiqing, @liulch, Is it possible to obtain unit level standard errors when using matrix completion in a two-way format? Thus far, I have not been able to construct them unless using the parametric version of IFE, in which case est.ind is not null.

Also, thank you for building this package, it's wonderful!

Best, Eddie

xuyiqing commented 4 years ago

It's possible. We'll think about a solution. But please note that treatment effects at the individual level are not identified under this model.

On Sun, Sep 15, 2019 at 10:14 PM Eddie Landesberg notifications@github.com wrote:

Is it possible to obtain unit level standard errors when using matrix completion in a two-way format? Thus far, I have not been able to construct them unless using the parametric version of IFE, in which case est.ind is not null.

Also, thank you for building this package, it's wonderful!

Best, Eddie

— You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub https://github.com/xuyiqing/gsynth/issues/28?email_source=notifications&email_token=AB2PKGD2T2Y66DEFFJGIY6LQJ4I3NA5CNFSM4IW5ZFW2YY3PNVWWK3TUL52HS4DFUVEXG43VMWVGG33NNVSW45C7NFSM4HLPR3SA, or mute the thread https://github.com/notifications/unsubscribe-auth/AB2PKGCYLYN53EF7KAYF2MDQJ4I3NANCNFSM4IW5ZFWQ .

-- Yiqing Xu

Assistant Professor Department of Political Science Stanford University http://yiqingxu.org/

elandesberg commented 4 years ago

Thank you for your prompt response! I didn't realize that these effects weren't identified with the matrix completion method - would be interested in understanding why if you'd be willing to point me to a relevant section of the literature. To confirm, these effects individual are identified under the IFE models, right? Which is why est.ind exists in the package when running them with inference = 'parametric'? cc: @Zaw5009

elandesberg commented 4 years ago

@xuyiqing Additional question, when you get a chance. If you wanted to obtain individual effects of each treatment unit (e.g., 14 geographies) using matrix completion, would you simply run the matrix completion algorithm in gsynth 14 times? Further, would you expect the average of these individual ATTs to approximately equal the ATT obtained with a single model fit (under which, as you mentioned, the individual effects are not identified)? Would you expect the precision to differ?