A lower 95% confidence bound θ̂ for a mean μ should satisfy P(μ < θ̂) = 0.05. Section 3.5 states that the Normal, basic pivotal and percentile bounds get this right only up to an error of order 1/√n (first-order accurate), while the studentized bound is off by order 1/n (second-order accurate). The reason is the Edgeworth expansion (3.23): the bootstrap copies the skewness term of the studentized pivot, but not of the unstudentized one, whose scale it has to estimate. Panel (a) simulates the miss rates for several n; panel (b) shows the 1/√n skewness correction for the chosen population.
Each data set r comes from its own seeded generator (seed + 7919 r) and its B resamples continue that stream. Bounds at level 0.95 for the mean: Normal X̄ − 1.645 seboot; basic pivotal 2X̄ − X̄*(0.95); percentile X̄*(0.05); studentized X̄ − z*0.95 σ̂/√n, with Z* = (X̄* − X̄)/(σ̂*/√n) (divisor n in σ̂). The grey band is 0.05 ± 2 Monte Carlo standard errors. In panel (b), the exact curve uses the Gamma law of the mean of exponential and χ²1 data; γ is the skewness of one observation.
code/ch03_bootstrap_jackknife.py (one_sided_bounds, os_demo, edgeworth_mean, edgeworth_t).