lsun20 / EventStudyInteract

MIT License
34 stars 12 forks source link

Is it possible to specify the parameter restriction not at t-1? #7

Open Alalalalaki opened 1 year ago

Alalalalaki commented 1 year ago

Thank you very much for building the package.

My question is that while the coefficient estimated in t-1 period is normalized to 0, I want to try some other specifications like t-2 or t-3 as the benchmark, or setting the average of the entire pre-treatment period as 0 as suggested by the recent survey paper of Miller 2023 JEP. I didn't find such setting information in the help page.

Is this kind of specification possible with EventStudyInteract? Thanks in advance.

lsun20 commented 1 year ago

Thank you very much for your interest in the package! The normalization is based on which pre-treatment period(s) is left out in {rel_time_list} so there is no default like t-1. Hope this helps!

Alalalalaki commented 1 year ago

Thank you very much for the very prompt reply! I see the point. Yes, the lack of t-1 coefficient is completely due to my construction of the estimation variables. I include all pre-treatment time variables except t-1 (F1event) in my case.

However, what if I want to set a constraint like constraint define 1 F1event + F2event + F3event = 0 and then include all the pre-treatment variables? Is this still possible with EventStudyInteract? The point is that in my estimated results I observe a dip at t-1 and I have learned from Miller 2023 JEP that using a joint constraint could reduce the standard deviation for pre-trend coefficients.

lsun20 commented 1 year ago

If the joint constraint you have in mind is constraint define 1 F2event =0 F3event = 0 then excluding F2event F3event effectively imposes the constraint. Perhaps this is what you had in mind?

Alalalalaki commented 1 year ago

Thanks for the reply! However, this is a little bit different from what I have in mind. If I understand it correctly, excluding F2event F3event adds two parameter restrictions and mute both coefficients at t-2 and t-3. Rather I want to hold a joint constraint of the average of t-1, t-2, and t-3 to be 0 on average. This helps to show all the pre-trend coefficients and to smooth the noise of a particular period.

lsun20 commented 1 year ago

I see! Operationally I think you can exclude F1event and include as regressors F2event-F1event and F3event-F1event to meet the constraint that their coefficients add up to zero.