I have a problem where I think subsampling is more appropriate than the bootstrap. (Reason in another post.)

However, I found no quick reference on subsampling CIs, and my naive inversion of the basic theory led to something that looks like the analogue of the "basic bootstrap" CI which seems to be used less often itself. Also, another question got a reply that subsampling applications usually "rely on asymptotic normality of the estimate", and "a bootstrap of some sort is more appropriate eg for asymmetric confidence intervals."

My (limited) understanding of subsampling does not seem to imply that.

I think I can get CIs as $$(\theta -\frac{\sqrt{b}}{\sqrt{n}}(\theta^{*}_{(1-\alpha)}-\theta);\theta -\frac{\sqrt{b}}{\sqrt{n}}(\theta^{*}_{(\alpha)}-\theta))$$ where $\theta^{*}_{(1-\alpha)}$ denotes the $1-\alpha$ percentile of the subsampled coefficients $\theta^{*}$ and $\theta$ is original estimate (so yes, everything could use $\hat{\theta}$ here).

So, if I have a nonparametric estimator for the "excess mass" at some parts of the distribution (implicitly using kernel density estimates usually unfit for the bootstrap), and I "simply" repeat the process with subsamples without replacement, how do I get a CI from the distribution of estimates? Take the raw percentiles (á la "percentile bootstrap") or invert the statistic from the general theory ("basic bootstrap").

Monte Carlos would be prohibitively expensive to compare these.

By the way, I know subsampling is not really a smart move without deriving something about the rate of convergence (simple square root above) and the right method to pick the size of the subsample (b). But is a "percentile bootstrap" any smarter in this case?

Finally, maybe a question worthy of its own posting, but if "b=n" is not required for the common bootstrap either (nor is no-replacement, thus subsampling could be a subset of bootstrap methods), why isn't the $\frac{\sqrt{b}}{\sqrt{n}}$ correction part of the bootstrap CI formulas (basic or BCa)?

  • 1
    $\begingroup$ I posted a similar question regarding subsampling and CIs. My approach mirrors yours I believe. stats.stackexchange.com/questions/363703/… $\endgroup$ Commented Aug 25, 2018 at 3:37
  • $\begingroup$ @compbiostats the above mentioned link is now dead, do you know why? $\endgroup$
    – Yair Daon
    Commented Oct 26, 2022 at 16:55
  • $\begingroup$ @YairDaon Hmmm... I don't know. I believe the question concerned deriving a confidence interval based on quantiles of an empirical distribution of a complicated ratio statistic where the estimated value of the parameter of interest fell outside the CI. So if $\hat{\theta}$ = 5 and I get a 95% CI of (3, 4), something's wrong. Perhaps one of the the CV moderators can chime in as to reason behind the broken link? $\endgroup$ Commented Oct 27, 2022 at 17:38
  • 1
    $\begingroup$ @YairDaon Stack Exchange has network-wide scripts that will delete questions under certain conditions. This script is colloquially called the "Roomba." meta.stackexchange.com/help/roomba $\endgroup$
    – Sycorax
    Commented Oct 28, 2022 at 12:44


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.