关于「分店都赢、合并却输:二乘二乘 k 表的反转条件」,《辛普森悖论与后门调整》给出的定量结论是什么?

结论: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。

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: 0.050, 0.304, 0.387, 0.430, 0.445

campus 1973 men applied 8442 admitted 44 %
campus 1973 women applied 4321 admitted 35 %
dept men_n men_rate wom_n wom_rate wom-men
A825 0.621 108 0.824 +0.203
B560 0.630 25 0.680 +0.050
C325 0.369 593 0.341 -0.029
D417 0.331 375 0.349 +0.018
E191 0.277 393 0.239 -0.038
F373 0.059 341 0.070 +0.011
pooled six: men 1198/26910.445 women 557/1835 = 0.304
departments where women's rate is higher4 of 6
adjusted to applicant mix (n4526): men 0.387 women 0.430
pooled gap -0.142 adjusted gap +0.043
share applying to A or B: men 0.515 women 0.072
查看完整验算程序
# Chapter 1. p11: Berkeley 1973, six largest departments, by sex.
# Data: campus totals from Bickel, Hammel and O'Connell (1975, Science);
# the six-department counts are the subset reproduced in textbooks
# (Freedman, Pisani and Purves) and in the UCBAdmissions data set.
# Exact rates with Fraction; the adjusted rates standardise both sexes
# to the pooled applicant mix of the six departments (back-door form).
from fractions import Fraction as F

CAMPUS = {"men": (8442, 44), "women": (4321, 35)}   # applicants, % adm.
# dept: men admitted, men applied, women admitted, women applied
DEPTS = [("A", 512, 825, 89, 108), ("B", 353, 560, 17, 25),
         ("C", 120, 325, 202, 593), ("D", 138, 417, 131, 375),
         ("E", 53, 191, 94, 393), ("F", 22, 373, 24, 341)]


def main():
    for k in ("men", "women"):
        n, pct = CAMPUS[k]
        print("campus 1973 {:5s} applied {}  admitted {} %".format(
            k, n, pct))
    print("dept  men_n  men_rate  wom_n  wom_rate  wom-men")
    ma = mn = wa = wn = 0
    women_ahead = 0
    for d, a1, n1, a2, n2 in DEPTS:
        r1, r2 = F(a1, n1), F(a2, n2)
        women_ahead = women_ahead + (1 if r2 > r1 else 0)
        print("{:>4s} {:6d} {:9.3f} {:6d} {:9.3f} {:+8.3f}".format(
            d, n1, float(r1), n2, float(r2), float(r2 - r1)))
        ma, mn, wa, wn = ma + a1, mn + n1, wa + a2, wn + n2
    pm, pw = F(ma, mn), F(wa, wn)
    print("pooled six: men {}/{} = {:.3f}  women {}/{} = {:.3f}".format(
        ma, mn, float(pm), wa, wn, float(pw)))
    tot = mn + wn
    adj_m = F(0)
    adj_w = F(0)
    for d, a1, n1, a2, n2 in DEPTS:
        w = F(n1 + n2, tot)
        adj_m = adj_m + w * F(a1, n1)
        adj_w = adj_w + w * F(a2, n2)
    print("departments where women's rate is higher: {} of {}".format(
        women_ahead, len(DEPTS)))
    print("adjusted to applicant mix (n = {}): men {:.3f} women {:.3f}"
          .format(tot, float(adj_m), float(adj_w)))
    print("pooled gap {:+.3f}   adjusted gap {:+.3f}".format(
        float(pw - pm), float(adj_w - adj_m)))
    sm = F(DEPTS[0][2] + DEPTS[1][2], mn)
    sw = F(DEPTS[0][4] + DEPTS[1][4], wn)
    print("share applying to A or B: men {:.3f} women {:.3f}".format(
        float(sm), float(sw)))


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 1 章「分店都赢、合并却输:二乘二乘 k 表的反转条件」(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