关于「技能与运气的方差分解:重复测量、组内相关与信度」,《均值回归与封面诅咒》给出的定量结论是什么?
结论:两年之间的相关系数就是技能方差在总方差中所占的份额,可以用单因素随机效应方差分析从重复测量中估出来;但它的估计本身并不精确,60 家门店、4 年数据给出的区间仍然很宽。斯皮尔曼—布朗公式告诉我们需要多少期数据才能让排名有意义,对单期信度低的指标,答案往往是几十期。二项抽样下的转化率需要数百乃至上千个订单才能分开技能与运气。技能确实存在,有时还很持久;方差分解的用处是把这句话变成一个可以算、可以检验的数。
数据口径:书中门店、达人、顾客与报告的参数均为示意;全部计算由书中程序生成;本书是非个人化的经营分析,不构成经营、法律或理财建议,涉及运程报告与原典之处只分析结构与概率,不作任何个人占断。 书中以标注框区分已发表的证据、可运行的 Python 计算(22 段程序)、原典与作者的假说。
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
| true: var_skill 0.360 var_luck 1.000 ICC 0.2647 | |
| one study: var_skill 0.417 var_luck 1.169 ICC 0.2630 | |
| 500 repeat studies: ICC 2.5% 0.121 median 0.269 | 97.5% 0.397 |
| reliability of a 4-year average | 0.5902 |
| pull a 4-year average toward the chain mean by | 0.4098 |
核心公式
\[y_{it} = \mu + s_i + e_{it},\]
查看完整验算程序
# Chapter 3. p31: Splitting a chain's spread into skill and luck.
# 60 stores observed for 4 years. score = mu + skill + luck with skill sd
# 0.6 and luck sd 1.0 (true single-year reliability 0.36 / 1.36). One-way
# random-effects ANOVA recovers the two variances from the data alone:
# var_skill = (MSB - MSW) / n, var_luck = MSW,
# ICC = var_skill / (var_skill + var_luck).
# We repeat the 4-year study 500 times to show how unsure the estimate is.
# Fixed seed; parameters illustrative.
# 单因素随机效应方差分析:从重复测量中分离技能方差与运气方差。
import math
import random
rng = random.Random(31)
STORES, YEARS = 60, 4
SK, LU = 0.6, 1.0
def study():
data = []
for _ in range(STORES):
s = rng.gauss(0.0, SK)
data.append([10.0 + s + rng.gauss(0.0, LU) for _ in range(YEARS)])
grand = math.fsum(math.fsum(r) for r in data) / (STORES * YEARS)
means = [math.fsum(r) / YEARS for r in data]
ssb = YEARS * math.fsum((m - grand) ** 2 for m in means)
ssw = math.fsum(math.fsum((v - m) ** 2 for v in r)
for r, m in zip(data, means))
msb = ssb / (STORES - 1)
msw = ssw / (STORES * (YEARS - 1))
vs = max(0.0, (msb - msw) / YEARS)
return vs, msw, vs / (vs + msw)
if __name__ == "__main__":
true_icc = SK ** 2 / (SK ** 2 + LU ** 2)
vs, vl, icc = study()
print("true: var_skill {:.3f} var_luck {:.3f} ICC {:.4f}"
.format(SK ** 2, LU ** 2, true_icc))
print("one study: var_skill {:.3f} var_luck {:.3f} ICC {:.4f}"
.format(vs, vl, icc))
iccs = sorted(study()[2] for _ in range(500))
print("500 repeat studies: ICC 2.5% {:.3f} median {:.3f} 97.5% {:.3f}"
.format(iccs[12], iccs[250], iccs[487]))
four = YEARS * true_icc / (1 + (YEARS - 1) * true_icc)
print("reliability of a {}-year average: {:.4f}".format(YEARS, four))
print("pull a 4-year average toward the chain mean by: {:.4f}"
.format(1.0 - four))唯一数据源与引用
范雅福(Fan, Y.). (2026). 《均值回归与封面诅咒、运程报告后验校准与塞翁失马序列独立性母典》第 3 章「技能与运气的方差分解:重复测量、组内相关与信度」(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-202 · v1.0.0 · CC-BY-NC-ND-4.0
BibTeX
@book{fan2026202,
author = {Fan, Yafu},
title = {Regression to the Mean and the Cover Jinx: Posterior Calibration of Fortune Reports and Serial Independence in the Tale of the Lost Horse},
year = {2026},
publisher = {5D Library},
edition = {1.0.0},
note = {DOI pending}
}