Wow. What do you do next? Nothing? I conjecture they might have a harder time recruiting reviewers in the future…

plus the paper isn't even retracted. :/

Nice figure!

Good idea - do you have a good example?

require more scrutiny of mechanisme and possible explanations for why the treatment effects can be so large. Doesn't always feel fair that either, but better than any alternative I've found.

my thoughts exactly. It's hard be skeptical as a referee thought without hanging ut on some specific methodological shorthcoming or explanation, without outright accusing the authors of fraud? When I find implausibly large effect sizes, but has no methods issues to hang it on, my go-to comment is to

www.statalist.org/forums/forum...

But it seems to work when my command sets e(cmdname) instead…

I don’t know exactly why, but if your program sets e(cmd), bootstrap will throw an error after the final bootstrap repetition has ended, during the display of the results. Had to simply remove setting e(cmd). Without e(cmd), however, you cannot use estimates store on the results.

Should have started before!

It gets better!

Sample splitting would work, I guess? Test in half, run either TWFE or saturated estimator in other half, reverse roles?

This is a clever idea! Very straightforward to implement. Is there an issue with pre-testing here, e.g. if researchers use this test, and run with TWFE if p>0.05? Similar in spirit to first stage testing with IV.

Mandatory pic of outdoors opportunities just 20 minutes from the city centre with the metro: