I don't have a Stata question as much as I do a conceptual statistics question. I'm writing an event-study synthetic control estimator that's good in situations where more than one unit is treated at different times, or staggered implementation.

While writing the syntax, I thought a lot about the recent advancements in difference-in-differences, namely the idea of how we might construct the donor pool/comparison group for newly treated units. The conversation in the DD literature seems to boil down to two main ideas: we can use the not-yet-treated units as the donor pool or never treatedunits.

I can see two appealing arguments for either case: If we use the never-treated units as our comparison group, we arguably are using the cleanest form of donors possible. Since they were never exposed to the intervention, assuming SUTVA and others, we can approximate an unbiased ATT by using the never-treated units since there's no chance of contamination of treatment effects by using already-treated units.

However, I also see a case for using the not-yet-treated units as donor units, arguably a better one: perhaps the ever-treated units may share latent similarities to each other on unobserved background traits, thereby allowing us to better model the real data generating process of the pre-intervention outcomes based on this fact.

Does anyone have any thoughts on how we might marry together these ideas in synthetic controls and differedifferences-in-differences?