关于「分层估计的抽样误差:区间、检验力与稀疏格子」,《辛普森悖论与后门调整》给出的定量结论是什么?

结论:要以 80 % 的检验力看出 0.05 的效应,每店每方案需要 393.5 人,向上取整为 394 人,三店两方案合计 2364 人。表格给出了检验力随样本量的变化:每店每方案 100 人时只有 0.292,200 人时 0.515,400 人时 0.806,600 人时 0.933,800 人时 0.979,1200 人时 0.998。模拟出的 400 人时的检验力是 0.814,与闭式的 0.806 相差在模拟误差之内。

数据口径:书中门店、方案与成交数均为示意;全部计算由书中程序生成。 第一章在伯克利六个最大院系上复算反转:合并后男性录取率 0.445、女性 0.304,按申请者构成标准化后男性 0.387、女性 0.430;三店示意报表中甲在每店都高 0.050,合并后却落后 0.2493,并给出反转可能的充要条件。

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: 0.292, 0.515, 0.806, 0.814, 0.933, 0.979

n per plan per branch for 80 % power393.5 -> 394
customers in total2364
n se power
100 0.03540.292
200 0.02500.515
400 0.01770.806
600 0.01450.933
800 0.01250.979
1200 0.01020.998
simulated power at n400 (1000 runs): 0.814
查看完整验算程序
# Chapter 6. p62: power of the branch-standardised gap.
# Rates from p12 (A beats B by 0.05 in each branch), branch weights
# w = (0.4, 0.3, 0.3); n customers per plan per branch. Normal
# approximation: power = Phi(d / se - z), se^2 = sum_s w_s^2 V_s / n,
# V_s = pA qA + pB qB. A seeded simulation checks one row.
import math
import random
from statistics import NormalDist

RA = [0.80, 0.45, 0.20]
RB = [0.75, 0.40, 0.15]
W = [0.4, 0.3, 0.3]
D = 0.05
ND = NormalDist()
ZA = ND.inv_cdf(0.975)


def se(n):
    v = [RA[s] * (1 - RA[s]) + RB[s] * (1 - RB[s]) for s in range(3)]
    return math.sqrt(math.fsum(W[s] ** 2 * v[s] for s in range(3)) / n)


def power(n):
    return ND.cdf(D / se(n) - ZA)


def simulate(n, reps, rng):
    hit = 0
    for _ in range(reps):
        d, v = [], []
        for s in range(3):
            a = len([1 for _ in range(n) if rng.random() < RA[s]]) / n
            b = len([1 for _ in range(n) if rng.random() < RB[s]]) / n
            d.append(W[s] * (a - b))
            v.append(W[s] ** 2 * (a * (1 - a) + b * (1 - b)) / n)
        if math.fsum(d) / math.sqrt(math.fsum(v)) > ZA:
            hit = hit + 1
    return hit / reps


def main():
    zb = ND.inv_cdf(0.8)
    v = [RA[s] * (1 - RA[s]) + RB[s] * (1 - RB[s]) for s in range(3)]
    need = (ZA + zb) ** 2 * math.fsum(
        W[s] ** 2 * v[s] for s in range(3)) / D ** 2
    print("n per plan per branch for 80 % power: {:.1f} -> {}".format(
        need, math.ceil(need)))
    print("customers in total: {}".format(6 * math.ceil(need)))
    print("   n   se      power")
    for n in (100, 200, 400, 600, 800, 1200):
        print("{:5d} {:.4f} {:8.3f}".format(n, se(n), power(n)))
    rng = random.Random(20261113)
    print("simulated power at n = 400 (1000 runs): {:.3f}".format(
        simulate(400, 1000, rng)))


if __name__ == "__main__":
    main()

唯一数据源与引用

范雅福(Fan, Y.). (2026). 《辛普森悖论与后门调整准则、多分店数据合并结论反转与因果图去混杂母典》第 6 章「分层估计的抽样误差:区间、检验力与稀疏格子」(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-302 · v1.0.0 · CC-BY-NC-ND-4.0

BibTeX
@book{fan2026302,
  author    = {Fan, Yafu},
  title     = {Simpson's Paradox and the Back-Door Criterion: Reversals When Branch Data Are Pooled and De-Confounding with Causal Graphs},
  year      = {2026},
  publisher = {5D Library},
  edition   = {1.0.0},
  note      = {DOI pending}
}

本书其他章节结论