Interesting new paper: arxiv.org/pdf/2503.09907 improves on both rdrobust and rdhonest! Quite compelling... 1/n - ThreadSky

paulgp.com • 19 days ago

Interesting new paper: https://arxiv.org/pdf/2503.09907

improves on both rdrobust and rdhonest!

Quite compelling... 1/n

Comments

paulgp.com•19 days ago

[improves on rdrobust and rdhonest by making some assumptions, I should say!]

paulgp.com•19 days ago

A killer quote: "A first observation is that rdrobust systematically undercovers the true regression discontinuity parameter. The basic intervals only [work] in 1 out of 9 settings. The [robust] intervals are wider, but still only achieve nominal coverage in 6 out
of the 9 settings."

paulgp.com•19 days ago

Here's how those methods change the CI for those benchmark studies as well:

paulgp.com•19 days ago

Then they propose a new estimator, which I need to understand, but the gains come from putting functional form restrictions on the shape of how the running variable affects the treatment effect.

paulgp.com•19 days ago

Two questions I would ask if the authors were in front of me:

1) How would this approach fair when compared in the setting proposed by Korting et al. (QJE, 2023): https://arxiv.org/abs/2112.03096 In this setting, the rdhonest setting does extremely well, and I always took that as evidence in its favor.

paulgp.com•19 days ago

2. Is this approach valid when the running variable is discrete (as in Imbens and Wager (2019) and rdhonest (Kolesar and Rothe))? That would be very helpful!

jiafengkevinchen.bsky.social•18 days ago

If I understand correctly, it’s doing the same conceptual
thing as rdhonest but imposing that CATE(x) is linear. If so, then probably can extend to discrete running variables (as long as CATE(x) being linear in x still makes sense)

paulgp.com•17 days ago

Yes, i think that’s right (although after emailing with Stefan, it sounds like there’s a subtlety in how the smoothness bound is chosen (which assumes data everywhere)).

donskerclass.bsky.social•19 days ago

Cute idea. Imposing additional regularity on the difference in means has been shown to improve minimax rates in @edwardhkennedy.bsky.social et al https://arxiv.org/abs/2203.00837 not in an RD setting or for inference. It makes me think similar techniques might apply beyond the exactly linear CATE case.

paulgp.com•19 days ago

oh neat, great connection.

oleftra.bsky.social•19 days ago

Yep. I was just talking about this the other day 😉

Comments

Posting Rules

Reply