The science
What is modelled
You hold one sample of n numbers drawn from a population you cannot see. The bootstrap treats the sample as a stand-in for the population: it draws n values with replacement from the sample (a "resample"), computes the statistic, and repeats B times. The spread of those B values estimates how much the statistic would vary from sample to sample, with no formula for its distribution.
Equations
- Resample:
x*i = xJ(i),J(i)uniform on {1..n}. Statisticθ* = T(x*)(mean or median). - Bootstrap standard error:
SEboot = sd(θ*1..θ*B). - Percentile interval: the
α/2and1-α/2quantiles of the sortedθ*values (linear interpolation between order statistics). - Closed form for the mean: when T is the mean, the exact (B→∞) bootstrap SE is
sqrt(Σ(xi - x̄)²) / n=s/√n · √((n-1)/n), the textbooks/√nwith the plug-in (divide by n) variance. Monte Carlo noise inSEbootis about1/√(2B)relative.
Checks run in this page
- SE check: 4000 fresh resamples (separate random stream) must reproduce the closed form within 3.5 standard errors of the Monte Carlo estimate (width computed from the measured kurtosis of the resampled means).
- Coverage check: 400 new samples are drawn from the known population, a percentile interval is built from each (400 resamples), and the fraction that contains the true value is compared with the nominal level, with a binomial standard error. For the mean, the exact Student-t interval is run on the same kind of samples as an independent reference (its critical value comes from an incomplete-beta bisection, not from the bootstrap).
Honest simplifications
- The percentile interval is the simplest bootstrap interval. It tends to under-cover for small n and for skewed data (try "Tiny n=8" or "Skewed, n=15": coverage lands visibly below nominal). BCa or bootstrap-t intervals correct part of that; they are not implemented here.
- The bootstrap cannot invent information that the sample does not contain: with n=8 the sample may miss the tail entirely, and no number of resamples fixes that.
- The median of a small sample takes few distinct values, so its bootstrap distribution is lumpy; this is real behaviour, not a bug.
- Populations are made-up textbook distributions with fixed parameters; coverage numbers carry Monte Carlo error (shown) and are not universal constants.
- Pseudo-random numbers come from a seeded generator, so a preset is exactly repeatable.