关于「辛普森反转:脱离环境的单句规则为何会指错方向」,《条件参数化规则蒸馏与在线校准》给出的定量结论是什么?

结论:第一章量化单句规则的代价,在示意决策表上「一律用 B」对 0.5433 的流量是错的。

数据口径:书中参数为示意,全部计算由书中程序生成。 本书是非个人化经营分析,不构成投资、法律或医疗意见,不对应任何真实客户或成交。

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. admission rate by department (men vs women)
  dept  men n  men rate  women n  women rate  women - men
   A     825   0.621      108     0.824      +0.203
   B     560   0.630       25     0.680      +0.050
   C     325   0.369      593     0.341      -0.029
   D     417   0.331      375     0.349      +0.018
   E     191   0.277      393     0.239      -0.038
   F     373   0.059      341     0.070      +0.011
B. pooled over the six departments
  men   2,691 applied, 1,198 admitted, rate 0.445
  women 1,835 applied, 557 admitted, rate 0.304
  departments where women do better: 4 of 6
C. odds ratio of admission, women vs men
  crude (pooled) 0.543   Mantel-Haenszel 1.105
D. share applying to the four harder departments C to F
  men 0.485   women 0.928
查看完整验算程序
# Chapter 1. p11: Simpson reversal in the Berkeley 1973 admissions table.
# Data: the six largest departments as published in the R dataset
# UCBAdmissions (Bickel, Hammel and O'Connell 1975). Public counts.
# For each department: men admitted, men rejected, women admitted,
# women rejected. We compare the pooled rates with the stratified ones
# and with the Mantel-Haenszel common odds ratio.
import math

DATA = [("A", 512, 313, 89, 19), ("B", 353, 207, 17, 8),
        ("C", 120, 205, 202, 391), ("D", 138, 279, 131, 244),
        ("E", 53, 138, 94, 299), ("F", 22, 351, 24, 317)]


def rate(a, r):
    return a / (a + r)


def main():
    print("A. admission rate by department (men vs women)")
    print("  dept  men n  men rate  women n  women rate  women - men")
    favour = 0
    for d, ma, mr, wa, wr in DATA:
        diff = rate(wa, wr) - rate(ma, mr)
        favour += 1 if diff > 0 else 0
        print("   {}   {:>5,}   {:.3f}    {:>5,}     {:.3f}      {:+.3f}"
              .format(d, ma + mr, rate(ma, mr), wa + wr, rate(wa, wr),
                      diff))
    men_a = int(math.fsum(r[1] for r in DATA))
    men_n = int(math.fsum(r[1] + r[2] for r in DATA))
    wom_a = int(math.fsum(r[3] for r in DATA))
    wom_n = int(math.fsum(r[3] + r[4] for r in DATA))
    print("B. pooled over the six departments")
    print("  men   {:,} applied, {:,} admitted, rate {:.3f}".format(
        men_n, men_a, men_a / men_n))
    print("  women {:,} applied, {:,} admitted, rate {:.3f}".format(
        wom_n, wom_a, wom_a / wom_n))
    print("  departments where women do better: {} of 6".format(favour))
    # odds ratio women : men, crude and Mantel-Haenszel
    crude = (wom_a * (men_n - men_a)) / ((wom_n - wom_a) * men_a)
    num = math.fsum(wa * mr / (ma + mr + wa + wr)
                    for _, ma, mr, wa, wr in DATA)
    den = math.fsum(wr * ma / (ma + mr + wa + wr)
                    for _, ma, mr, wa, wr in DATA)
    print("C. odds ratio of admission, women vs men")
    print("  crude (pooled) {:.3f}   Mantel-Haenszel {:.3f}".format(
        crude, num / den))
    hard = ("C", "D", "E", "F")
    m_h = int(math.fsum(r[1] + r[2] for r in DATA if r[0] in hard))
    w_h = int(math.fsum(r[3] + r[4] for r in DATA if r[0] in hard))
    print("D. share applying to the four harder departments C to F")
    print("  men {:.3f}   women {:.3f}".format(m_h / men_n, w_h / wom_n))


if __name__ == "__main__":
    main()

唯一数据源与引用

范雅福(Fan, Y.). (2026). 《高阶知识蒸馏进化、条件参数化编译与实盘转化反馈校准母典》第 1 章「辛普森反转:脱离环境的单句规则为何会指错方向」(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-377 · v1.0.0 · CC-BY-NC-ND-4.0

BibTeX
@book{fan2026377,
  author    = {Fan, Yafu},
  title     = {Next-Generation Knowledge Distillation, Context-Parametric Compilation and Live-Conversion Feedback Calibration},
  year      = {2026},
  publisher = {5D Library},
  edition   = {1.0.0},
  note      = {DOI pending}
}

本书其他章节结论