关于「东西方原典考辨:《吕氏春秋·察传》与 Yule 1903」,《辛普森悖论与后门调整》给出的定量结论是什么?

结论:第七章考辨《吕氏春秋·察传》「辞多类非而是,多类是而非」与 Yule 1903。

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

(a) flip chance per retelling q0.1
after1 retellings P(wrong) = 0.1000
after2 retellings P(wrong) = 0.1800
after3 retellings P(wrong) = 0.2440
after5 retellings P(wrong) = 0.3362
after 10 retellings P(wrong)0.4463
after 20 retellings P(wrong)0.4942
(b) h looks-right-is-wrong looks-wrong-is-right total
0.100.0600 0.0585 0.1185
0.250.1244 0.1183 0.2427
0.400.1559 0.1553 0.3113
0.490.1736 0.1747 0.3483
查看完整验算程序
# Chapter 7. p71: "white becomes black after several retellings" and
# "seems right but is wrong", written as two checkable quantities.
# (a) Each retelling flips a yes/no claim with probability q; after n
#     retellings P(wrong) = (1 - (1 - 2q)^n) / 2, exact Fraction.
# (b) Random three-branch worlds (no sampling noise): branch weights
#     uniform on the simplex, B rates uniform in [0.05, 0.85], branch
#     effects uniform in [-0.1, 0.1], P(A | branch) uniform in
#     [0.5 - h, 0.5 + h]. How often does the pooled gap carry the
#     wrong sign relative to the back-door average effect?
# Fixed seed; parameters illustrative.
import math
import random
from fractions import Fraction as F

WORLDS = 20000


def world(rng, h):
    g = [rng.expovariate(1.0) for _ in range(3)]
    t = math.fsum(g)
    pz = [x / t for x in g]
    pb = [rng.uniform(0.05, 0.85) for _ in range(3)]
    ef = [rng.uniform(-0.1, 0.1) for _ in range(3)]
    pa = [rng.uniform(0.5 - h, 0.5 + h) for _ in range(3)]
    ate = math.fsum(pz[s] * ef[s] for s in range(3))
    ma = math.fsum(pz[s] * pa[s] for s in range(3))
    ya = math.fsum(pz[s] * pa[s] * (pb[s] + ef[s]) for s in range(3))
    yb = math.fsum(pz[s] * (1 - pa[s]) * pb[s] for s in range(3))
    pooled = ya / ma - yb / (1 - ma)
    return ate, pooled


def main():
    q = F(1, 10)
    print("(a) flip chance per retelling q = {}".format(float(q)))
    for n in (1, 2, 3, 5, 10, 20):
        p = (1 - (1 - 2 * q) ** n) / 2
        print("    after {:2d} retellings P(wrong) = {:.4f}".format(
            n, float(p)))
    rng = random.Random(20261113)
    print("(b) h   looks-right-is-wrong  looks-wrong-is-right  total")
    for h in (0.1, 0.25, 0.4, 0.49):
        a = b = 0
        for _ in range(WORLDS):
            ate, pooled = world(rng, h)
            if pooled > 0 and ate < 0:
                a = a + 1
            elif pooled < 0 and ate > 0:
                b = b + 1
        print("  {:5.2f} {:15.4f} {:21.4f} {:12.4f}".format(
            h, a / WORLDS, b / WORLDS, (a + b) / WORLDS))


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 7 章「东西方原典考辨:《吕氏春秋·察传》与 Yule 1903」(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