Consider different advice about Standard Errors on Average Treatment effects

[jwbowers]

Putting this here so as not to forget it:

https://htmlpreview.github.io/?https://github.com/acoppock/Green-Lab-SOP/blob/master/Green_Lab_SOP.html#standard-errors-confidence-intervals-and-significance-tests

[benthestatistician]

I'll be happy to go with Lin et al's recommendation of Bell-McCaffrey (which seems to me a generalization/adaption of HC2 to the sampling context). I'm also open to variations on that recommendation. (My current sense of things is that there's a bit less at stake in what specific finite sample correction to use once one migrates from Wald-type tests & CIs over to robust score tests and CIs; but I digress.)

Ease of software implementation is important here. One component of that, in turn, is ease of addressing clustering, when present. In principle this is a simple adjustment to the meat calculation of a sandwich estimator, once that could be readily accommodated in the function signatures of sandwich::sandwich and sandwich::meat; but the authors of that package don't seem to have done that. @josherrickson has some nice work in this direction, here. I don't know what else is out there.

[jwbowers]

Winston/Don/Alex's code for BMlmSE() modified from https://github.com/kolesarm/Robust-Small-Sample-Standard-Errors, and the clubSandwich package https://github.com/jepusto/clubSandwich both also have some code for clustered assignment. I don't know about stratified clustered assignment (or what kinds of other design limits they have).

[benthestatistician]

Thanks much for those! clubSandwich in particular looks impressive. I don't think this will come up before @jwbowers's half of the course, so I don't think I'll have to make any decisions about which to recommend -- instead I'll just rest assured that we've got 'em covered when the time comes.