关于「随机效应的代价:异质性、预测区间与贝叶斯因子」,《自我损耗与饥饿法官》给出的定量结论是什么?

结论:随机效应模型承认研究之间的差异,但当小研究系统性偏大时,它会把偏差放大:示意数据中,固定效应为 0.140,随机效应为 0.383,大检验的权重从 0.694 降到 0.157。预测区间从 −0.078 到 0.844,比置信区间诚实得多。贝叶斯因子显示两次复制中等程度地支持零假设,但若事先强烈相信效应,后验概率仍有 0.5565。下一章离开统计方法,讨论四种关于「为什么会疲劳」的竞争解释,并找出能区分它们的实验设计。

数据口径:书中参数为示意;全部计算由书中程序生成;本书不是心理或医疗建议,书中流程与工具为非个人化经营建议。 本书把两项证据放到预注册复制、元分析偏差检验与案件排序模拟面前,看它们还剩多少。

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: 0.140, 0.157, 0.383, 0.694, 0.844

k 13, Q 30.49 on 12 df, tau^2 0.0475, I^2 0.606
fixed effect   d 0.140  95% CI 0.069 to 0.210
random effects d 0.383  95% CI 0.211 to 0.555
95% prediction interval (normal) -0.078 to 0.844
weight of the big replication: fixed 0.694, random 0.157
without the replication: fixed 0.366, random 0.398
studies with n <= 30 per group: 7 of 13
查看完整验算程序
# Chapter 4. p41: Random effects gives small studies a louder voice.
# Thirteen illustrative studies shaped like a depletion literature: many
# small two-group studies with large d, a few larger ones with small d,
# and one 23-lab-sized replication (n = 1,070 per group, d = 0.04).
# Fixed effect versus DerSimonian-Laird random effects: estimate, Q,
# tau^2, I^2, a normal-approximation 95% prediction interval, and the
# weight each model gives the replication. Data illustrative only.
# 随机效应模型把更多权重交给小研究:示意数据。
import math
from statistics import NormalDist

Z975 = NormalDist().inv_cdf(0.975)
STUDIES = [(0.95, 15), (0.80, 18), (0.72, 20), (0.66, 22), (0.58, 25),
           (0.61, 28), (0.45, 30), (0.52, 34), (0.38, 40), (0.30, 50),
           (0.21, 80), (0.10, 120), (0.04, 1070)]


def se_of(d, n):
    return math.sqrt(2 / n + d * d / (4 * n))


def pool(st):
    ds = [d for d, n in st]
    v = [se_of(d, n) ** 2 for d, n in st]
    w = [1 / x for x in v]
    sw = math.fsum(w)
    fe = math.fsum(a * b for a, b in zip(w, ds)) / sw
    q = math.fsum(a * (b - fe) ** 2 for a, b in zip(w, ds))
    k = len(ds)
    c = sw - math.fsum(a * a for a in w) / sw
    tau2 = max(0.0, (q - (k - 1)) / c)
    wr = [1 / (x + tau2) for x in v]
    swr = math.fsum(wr)
    re = math.fsum(a * b for a, b in zip(wr, ds)) / swr
    se_re = 1 / math.sqrt(swr)
    i2 = max(0.0, (q - (k - 1)) / q)
    return dict(fe=fe, se_fe=1 / math.sqrt(sw), re=re, se_re=se_re,
                q=q, tau2=tau2, i2=i2, wfe=w[-1] / sw, wre=wr[-1] / swr)


r = pool(STUDIES)
print("k {}, Q {:.2f} on {} df, tau^2 {:.4f}, I^2 {:.3f}".format(
    len(STUDIES), r["q"], len(STUDIES) - 1, r["tau2"], r["i2"]))
print("fixed effect   d {:.3f}  95% CI {:.3f} to {:.3f}".format(
    r["fe"], r["fe"] - Z975 * r["se_fe"], r["fe"] + Z975 * r["se_fe"]))
print("random effects d {:.3f}  95% CI {:.3f} to {:.3f}".format(
    r["re"], r["re"] - Z975 * r["se_re"], r["re"] + Z975 * r["se_re"]))
half = Z975 * math.sqrt(r["tau2"] + r["se_re"] ** 2)
print("95% prediction interval (normal) {:.3f} to {:.3f}".format(
    r["re"] - half, r["re"] + half))
print("weight of the big replication: fixed {:.3f}, random {:.3f}"
      .format(r["wfe"], r["wre"]))
small = pool(STUDIES[:-1])
print("without the replication: fixed {:.3f}, random {:.3f}".format(
    small["fe"], small["re"]))
nsmall = len([1 for d, n in STUDIES if n <= 30])
print("studies with n <= 30 per group: {} of {}".format(
    nsmall, len(STUDIES)))

唯一数据源与引用

范雅福(Fan, Y.). (2026). 《自我损耗大规模复制失败、法官饥饿效应排序伪影与决策疲劳日程母典》第 4 章「随机效应的代价:异质性、预测区间与贝叶斯因子」(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-196 · v1.0.0 · CC-BY-NC-ND-4.0

BibTeX
@book{fan2026196,
  author    = {Fan, Yafu},
  title     = {The Failed Replication of Ego Depletion, the Ordering Artefact behind the Hungry Judges and Decision-Fatigue Scheduling},
  year      = {2026},
  publisher = {5D Library},
  edition   = {1.0.0},
  note      = {DOI pending}
}

本书其他章节结论