Both methods estimate the standard error of a statistic Tn from one data set. The bootstrap recomputes Tn on resamples drawn with replacement (Section 8.2); the jackknife recomputes it on the n data sets with one observation left out and inflates their spread: vjack = ((n − 1)/n) Σ(T(−i) − T̄)² (Section 8.5.1). Here the population is known, so the true standard error is known too: panel (a) compares the three spreads for one data set, and panel (b) shows how far each estimate lands from the truth over R data sets.
Data set r comes from its own seeded generator (seed + 7919 r) and its B resamples continue that stream. The true standard error is the standard deviation of Tn over 4000 further data sets drawn with fixed seeds, so it does not change when you press the button. The "nerve" population is the 799 nerve-pulse waiting times of the book, sampled with replacement; its values are recorded to 0.01 s, so it has many ties. The median of an even number of values is the average of the two middle ones; quantiles interpolate linearly. In panel (a) the jackknife values are shown inflated, T̄ + √(n − 1)(T(−i) − T̄), so that their variance (divisor n) is exactly vjack.
code/ch08_bootstrap.py (jackknife, bootstrap, jk_demo; Figure 8.S5).