Accessing influence function values in aggte()

[PaoloZacchia]

Good day,

I hope someone can help me address a question

I am working on a project where I need to aggregate multiple ATT(g,t) estimates in a way not allowed by the package function aggte(), sometimes nonlinearly.

The output of function att_gt() allows to retrieve the values of the influence functions stored in a matrix (where if I am not mistaken, the row dimension matches the treated observations in the data, while the column dimension matches the number of estimated ATT(g,t) parameters). This is extremely useful.

However, the corresponding influence function stored by aggte() is one vector only. What is that? I can't figure this out. Is that related to the "grand" ATT that aggregates all the parameters? Or is that related to the term \xi from the original working paper version of the CS paper (in the discussion about the asymptotics and the bootstrap)?

I hope you can help me unravel this mystery!

Best,
Paolo

[pedrohcgs]

Hi Paolo,
Thanks for the question.

We do report two types of influence functions with aggte: on for the overall summary parameter, and one for the event-study or there other partially aggregated parameters.

For example, when you use did::aggte(type = "dynamic"), the term dynamic.inf.func.e inside inf.function is a N x E matrix, with E the number of event times considered. the dynamic.inf.func is the IF for the most-aggregated parameter.

Let us know if this helps.

Thanks

[PaoloZacchia]

Hello Pedro,

thanks a lot for your reply! Yesterday I could finally check this out, and I am happy to report that I found what I needed! I did not expect I had to dig deeper into the inf.function nest, my bad.

Just one clarification question about the E columns of the dynamic.inf.func.e matrix. Are they ordered from the past to the future relative to the treatment year? So that if, say, the event times run from -3 to +3, then E = 7 and the first column matches -3, the second column matches -2, and so on? I would like to make this sure.

Thanks a lot for the package and the dedication to address questions by users like me!

Best,
Paolo

[bcallaway11]

I am almost sure yes about the order. You could confirm by using the influence function directly to get the standard errors and comparing it to what the package gives.