Clustered standard errors

[catabia]

Don't think there is necessarily a bug, but I need clarification on an issue with the clustervar argument in att_gt. Docs say that you may enter a vector of up to two variables for clustering, as long as one of them is equivalent to idname. However, when I try to cluster on both state and idname (which is actually a column in my dataset that is also input as idname for the function--sorry for the confusing variable names), I get the following error:

I don't get an error when the only variable I choose is state. Is this because idname is already automatically included in the cluster?

[bcallaway11]

Yes, you are right, I think the issue is with the documentation. att_gt will automatically cluster by the "id" variable.

Brant

[catabia]

Hi Brant! Thanks so much for getting back to me quickly! I really appreciate how easy your package makes computing staggered DiDs. So to be absolutely sure that I understand this correctly, when I set clustervars='state', I am clustering on both idname and state? Not state instead of idname?

[pedrohcgs]

Hi Hannah, Since state “nests” idname, in this case you are effectively clustering at the state level. Thanks

On Thu, Apr 7, 2022 at 12:20 Hannah Catabia ***@***.***> wrote: Hi Brant! Thanks so much for getting back to me quickly! I really appreciate how easy your package makes computing staggered DiDs. So to be absolutely sure that I understand this correctly, when I set *clustervars='state'*, I am clustering on *both* *idname* and *state*? — Reply to this email directly, view it on GitHub <#123 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/ABE734563ZEFK6RZTYCKVPLVD4YQPANCNFSM5S2CO4FA> . You are receiving this because you are subscribed to this thread.Message ID: ***@***.***>

-- =============================== Pedro H. C. Sant'Anna Department of Economics Vanderbilt University 615-875-8448 (phone) ***@***.*** https://pedrohcgs.github.io ===============================

[catabia]

Thank you!

[catabia]

Another quick question. Say that the idname and the other chosen cluster variable are not nested. For instance, say I set clustervars=year. Since they are not nested, how is the cluster-robust variance matrix calculated? Is it this method:

That is to stay, adding the variance matrix for idname clusters to the variance matrix for year clusters, then subtracting the variance matrix for idname interacted with year?

[bcallaway11]

The way we calculate clustered standard errors (both in nested and non-nested cases) are by using the multiplier bootstrap and making the "same" draw of the weights for units in the same clusters. I am not sure if this is roughly analogous to the formula that you sent.

@pedrohcgs, any comments?