关于「门店小实验分析模板、置换检验器与证伪方案」,《门店小实验的重抽样推断》给出的定量结论是什么?

结论:第八章给出门店小实验分析模板、蒸馏小工具「置换检验器」与预注册证伪方案。书中以标注框区分已发表的证据、可运行的 Python 计算(20 段程序)、原典与作者的假说。

数据口径:书中参数为示意;全部计算由书中程序生成;本书属于非个人化经营建议。 书中以标注框区分已发表的证据、可运行的 Python 计算(20 段程序)、原典与作者的假说。

English version

对照数据表

下表数字由本章 Python 程序原样打印,可重新运行复现。 程序只用 Python 标准库与结论吻合的打印值: 20

new-layout days: 1 2 5 6 8 10 14
raw mean old 1013.1  new 1092.6  diff 79.4  lift 0.078
secondary, raw: exact p 1698 of 3432 = 0.4948
primary, centred: diff 79.4  exact p 20 of 3432 = 0.0058
95% percentile interval [40.8, 117.6]
95% BCa interval        [38.4, 115.2]
pre-registered decision at 0.05: adopt new layout
查看完整验算程序
# Chapter 8. p81: the shop trial analysis template, end to end.
# Pre-registered before day 1: 14 trading days; 7 drawn by lot (seed
# 20261126) for the new layout; outcome = daily takings in yuan.
# Primary statistic: difference in mean (new minus old) of takings
# centred on their own block mean, blocks fixed in advance (weekend =
# days 6, 7, 13, 14; weekdays = the rest). Centring does not depend on
# the assignment, so the permutation test stays exact. Two-sided exact
# test over all C(14, 7) = 3432 assignments at 0.05; no interim looks.
# Secondary: the raw difference. Always report a 95 percent two-sample
# bootstrap interval of the primary statistic (resampling within each
# arm, B = 4000, percentile and simplified BCa, leave-one-out
# jackknife). Takings are illustrative.
import itertools
import math
import random
from statistics import NormalDist

TAKINGS = [985, 1012, 905, 872, 1040, 1395, 1268,
           905, 958, 951, 921, 866, 1302, 1360]
WEEKEND = (5, 6, 12, 13)
N01 = NormalDist()


def mean(v):
    return math.fsum(v) / len(v)


def quantile(sv, q):
    pos = q * (len(sv) - 1)
    i = int(pos)
    if i >= len(sv) - 1:
        return sv[-1]
    return sv[i] + (pos - i) * (sv[i + 1] - sv[i])


def centred(y):
    wk = [y[i] for i in WEEKEND]
    wd = [y[i] for i in range(len(y)) if i not in WEEKEND]
    mk, md = mean(wk), mean(wd)
    return [y[i] - (mk if i in WEEKEND else md) for i in range(len(y))]


def exact_p(y, new):
    k, tot = len(new), math.fsum(y)
    s0 = math.fsum(y[i] for i in new)
    d0 = abs(s0 / k - (tot - s0) / (len(y) - k))
    ge = cnt = 0
    for idx in itertools.combinations(range(len(y)), k):
        s = math.fsum(y[i] for i in idx)
        cnt += 1
        ge += abs(s / k - (tot - s) / (len(y) - k)) >= d0 - 1e-9
    return ge, cnt


def boot_diff(a, b, nboot, rng):
    out = [mean(rng.choices(b, k=len(b))) - mean(rng.choices(a, k=len(a)))
           for _ in range(nboot)]
    return sorted(out)


def bca_diff(a, b, bd, alpha=0.05):
    th = mean(b) - mean(a)
    below = len([v for v in bd if v < th])
    z0 = N01.inv_cdf(below / len(bd))
    jack = ([mean(b) - mean(a[:i] + a[i + 1:]) for i in range(len(a))]
            + [mean(b[:i] + b[i + 1:]) - mean(a) for i in range(len(b))])
    jm = mean(jack)
    d = [jm - v for v in jack]
    acc = math.fsum(v ** 3 for v in d) / (
        6.0 * math.fsum(v ** 2 for v in d) ** 1.5)
    ends = []
    for q in (alpha / 2, 1 - alpha / 2):
        zq = N01.inv_cdf(q)
        ends.append(quantile(bd, N01.cdf(z0 + (z0 + zq)
                                         / (1 - acc * (z0 + zq)))))
    return ends


if __name__ == "__main__":
    rng = random.Random(20261126)
    new = sorted(d - 1 for d in rng.sample(range(1, 15), 7))
    old = [i for i in range(14) if i not in new]
    ra = [TAKINGS[i] for i in old]
    rb = [TAKINGS[i] for i in new]
    print("new-layout days: {}".format(" ".join(str(i + 1) for i in new)))
    print("raw mean old {:.1f}  new {:.1f}  diff {:.1f}  lift {:.3f}".format(
        mean(ra), mean(rb), mean(rb) - mean(ra), mean(rb) / mean(ra) - 1))
    ge0, cnt = exact_p(TAKINGS, new)
    print("secondary, raw: exact p {} of {} = {:.4f}".format(
        ge0, cnt, ge0 / cnt))
    adj = centred(TAKINGS)
    a = [adj[i] for i in old]
    b = [adj[i] for i in new]
    ge, cnt = exact_p(adj, new)
    print("primary, centred: diff {:.1f}  exact p {} of {} = {:.4f}".format(
        mean(b) - mean(a), ge, cnt, ge / cnt))
    bd = boot_diff(a, b, 4000, random.Random(20261127))
    print("95% percentile interval [{:.1f}, {:.1f}]".format(
        quantile(bd, 0.025), quantile(bd, 0.975)))
    lo, hi = bca_diff(a, b, bd)
    print("95% BCa interval        [{:.1f}, {:.1f}]".format(lo, hi))
    verdict = "adopt new layout" if ge / cnt <= 0.05 else "keep old layout"
    print("pre-registered decision at 0.05: {}".format(verdict))

唯一数据源与引用

范雅福(Fan, Y.). (2026). 《自助法与置换检验、小样本门店实验不靠正态假设的重抽样推断母典》第 8 章「门店小实验分析模板、置换检验器与证伪方案」(版本 1.0.0). 5D Library. DOI 待发布

DOI 待发布 · 专著页正式发布后上线(本站暂无专著页) · 5DL-B-308 · v1.0.0 · CC-BY-NC-ND-4.0

BibTeX
@book{fan2026308,
  author    = {Fan, Yafu},
  title     = {The Bootstrap and Permutation Tests: Resampling Inference for Small Shop Experiments without Normality Assumptions},
  year      = {2026},
  publisher = {5D Library},
  edition   = {1.0.0},
  note      = {DOI pending}
}

本书其他章节结论