vcovDA's applicable to paired assignment, w/ illustrative vignette

[benthestatistician]

• [x] Review econometric lit for advice re translation to "clusters" of pairs vs paired units
• [ ] Decide default design-based accommodation
• [x] Implementation
• [x] Brief end-user oriented vignette, demonstrating our semantics, explaining decisions being made implicitly and (as appropriate) illustrating options offered with our vcovDA calcs.

[benthestatistician]

(Xinhe and Josh, pls don't hesitate to update & add appropriate detail to issue statement.)

[benthestatistician]

The 2 papers I shared w/ @jwasserman2 today, as sources about the econometric conventional wisdom regarding sandwich estimation of covariance after paired assignment of clusters:

• cameronMiller15, cluster robust SEs for practice.pdf
• abadieSpiess21, Robust Post Matching Inference.pdf

[benthestatistician]

@jwasserman2 and I had a brief discussion today of the implementation of logic deciding which flavor of design-based variance estimator should be applied in which strata. This post is a followup to that discussion.

It would be useful for this purpose to have abstracted a table from the Design tallying the numbers of assignment unit by block, e.g.

	Condi	tion
Block	C	T
B1	1	3
B2	2	2

I promised I'd poke around for some related code elsewhere. Found it not to exist in the form that I had remembered. But I'll point out that flexida::treatment_table() is doing something related and is perhaps adaptible to this purpose.

[josherrickson]

Edit: Nevermind I was completely misremembering how that function worked. I do agree that a more generic table'ing function could be useful.

[josherrickson]

I expanded treatment_table into the new dtable() function (alias'd to design_table() for now but I'm considering removing that).

dtable(design, "treatment") replicates the existing functionality of treatment_table(), but now supports e.g. dtable(design, "t", "b") to obtain the table you drew above.

> data(simdata)
> des <- rct_design(z ~ uoa(cid1, cid2) + block(bid), data = simdata)
> dtable(des, "tr", "bl")
      treatment
blocks 0 1
     1 3 1
     2 2 1
     3 1 2

(The 2nd and 3rd arguments, x and optionally y, match against the full names "treatment", "blocks", etc, so x = "tre" for example is valid.)

[jwasserman2]

Thoughts on some of the implementation here, specifically the "cluster level" dispatch, have been updated by #85

[benthestatistician]

Please see list of outstanding to-do's for the vignette at top of thread for #179.