fwildclusterboot icon indicating copy to clipboard operation
fwildclusterboot copied to clipboard

Published 20 hours ago • verified publisher icon s3alfisc

Modified Jackknife

Open kylebutts opened this issue 2 months ago • 2 comments

https://users.ssc.wisc.edu/~bhansen/papers/tcauchy.pdf

Leaving this paper here as I might try to submit a PR to implement this modified version by Bruce Hansen. I think we might already be doing the generalized inverse he recommends, though

kylebutts avatar Sep 14 '25 23:09 kylebutts

Thanks, I'll take a closer look myself, haven't read the paper yet. Some of the "newer" bootstrap versions from MacKinnon et al are build on the jackknife and rely on the generalized inverse as far as I remember 🤔

s3alfisc avatar Sep 15 '25 19:09 s3alfisc

It is very similar and should be a very easy PR:

  1. No $(G - 1) / G$
  2. Generalized inverse for all clusters (so don't drop any non-invertible leave-g-out)
  3. Center based on full OLS and not average of leave-g-out Image

Theorem 1 shows that our recommended jackknife estimator $V_5$ is never downward biased, or conservative. This means that in any regression context, and any sample size, we can be confident that the jackknife estimator is not downward biased.

So this is a small refinement that works in a very broad set of categories. A no brainer win IMO

kylebutts avatar Sep 15 '25 19:09 kylebutts