AI ENGVisual Encyclopedia

MODULE 7: STATISTICAL RIGOR · SCENE 21

Bootstrap & Multiple Comparisons

Resample to build intervals without distributional assumptions — then pay the tax for every additional comparison you make.

α LEVELα = 0.05HOLM REJECTS · 2/4
1000 RESAMPLES · MEAN 0.692 · 95% INTERVAL [0.655, 0.727]FOUR SLICES TESTED — REPORTING THE BEST WITHOUT CORRECTION IS P-HACKINGp = 0.001 ✓ REJECTp = 0.01 ✓ REJECTp = 0.04 ✗ KEEPp = 0.2 ✗ KEEPTWENTY UNCORRECTED TESTS EXPECT ONE FALSE “DISCOVERY” — HOLM SPENDS α WHERE IT EARNS IT

The bootstrap frees you from normality assumptions; Holm frees you from fooling yourself across many slices.

TECHNICAL BREAKDOWNModule 7: Statistical Rigor

Resample for intervals, then pay for every comparison

The bootstrap builds a sampling distribution by resampling the observed data, without assuming normality — useful for medians, ratios, and Elo. But testing many slices at once inflates false positives. The multiple-comparisons tax (Bonferroni, Holm) must be paid on every additional hypothesis.

Bootstrap

Draw n items with replacement many times, recompute the statistic each time, and read percentiles as the interval.

Family-wise error

Twenty independent tests at α = 0.05 expect one false positive by chance alone.

Holm correction

Sort p-values and compare each to α/(m − i + 1), stopping at the first failure. Uniformly more powerful than Bonferroni.

MATHEMATICAL FORMULATION · HOLM STEP-DOWN
reject p_(i) if p_(i) ≤ α / (m − i + 1)

With p-values [0.001, 0.01, 0.04, 0.20] and α = 0.05: the first two fall below 0.0125 and 0.0167, the third (0.04) does not clear 0.025, so three comparisons are rejected in sequence — exactly two.

REAL-WORLD PRODUCTION ENGINEERING
  • Arena Elo intervals are frequently computed with the bootstrap over battles.
  • Benchmark papers comparing many models across many tasks must correct for multiplicity or report it explicitly.