5D Library › 塔罗超几何抽样、雷诺曼网格拓扑与卡巴拉生命树图论 · English

第六章 巴纳姆效应:覆盖模型与选择性记忆

福勒的课堂演示

1948 年,福勒给学生做了一份「人格测验」,随后每人收到一份「个人化」的描述,实际上所有人收到的都是同一份由星座书拼成的文字;学生们普遍给出很高的准确度评分 (Forer 1949)。此后大量研究证实,笼统、正面、似乎针对个人的描述会被普遍接受,这被称为巴纳姆效应或个人确认谬误 (Dickson and Kelly 1985)。本章用两个简单模型说明它为何必然出现在任何解读中。

覆盖模型

设一句陈述符合全部情境中的比例为 ff。听者用自己生活中 MM 个彼此独立的经历去对照,只要有一个对得上就接受它,接受的概率为 1−(1−f)M1 – (1-f)^M。

推导

# Chapter 6. p13: Why vague statements "fit" - a coverage model.
# A statement fits a fraction f of situations. A listener checks it
# against M independent episodes of their own life and accepts it if
# any one matches: P(accept) = 1 - (1 - f)^M. Vagueness and a rich
# memory together push acceptance towards 1. A statement that fits
# everyone equally cannot tell one person's reading from another's:
# in a choice among k readings the chance of picking one's own stays 1/k.
if __name__ == "__main__":
    fs = (0.02, 0.05, 0.10, 0.20, 0.50)
    Ms = (1, 5, 10, 30)
    print("   f \\ M  " + "".join("{:>8d}".format(m) for m in Ms))
    for f in fs:
        print("   {:5.2f}  ".format(f) + "".join(
            "{:8.3f}".format(1 - (1 - f) ** m) for m in Ms))
    f, M, k = 0.10, 10, 10
    p = 1 - (1 - f) ** M
    print("\n   a reading of {} statements with f = {:.2f}, M = {}:".format(
        k, f, M))
    print("   expected statements accepted: {:.2f} of {}".format(k * p, k))
    print("   the same reading given to anyone else: {:.2f} of {}".format(
        k * p, k))
    for choices in (2, 3, 4):
        print("   own reading picked out of {}: {:.4f} (chance)".format(
            choices, 1 / choices))
    for fs_ in ((0.10, 0.02), (0.30, 0.02)):
        own = 1 - (1 - fs_[0]) ** M
        other = 1 - (1 - fs_[1]) ** M
        print("   specific statement, fits owner {:.2f} / others {:.2f}:"
              " accept {:.3f} vs {:.3f}".format(fs_[0], fs_[1], own, other))
#    f \ M         1       5      10      30
#     0.02     0.020   0.096   0.183   0.455
#     0.05     0.050   0.226   0.401   0.785
#     0.10     0.100   0.410   0.651   0.958
#     0.20     0.200   0.672   0.893   0.999
#     0.50     0.500   0.969   0.999   1.000
#
#    a reading of 10 statements with f = 0.10, M = 10:
#    expected statements accepted: 6.51 of 10
#    the same reading given to anyone else: 6.51 of 10
#    own reading picked out of 2: 0.5000 (chance)
#    own reading picked out of 3: 0.3333 (chance)
#    own reading picked out of 4: 0.2500 (chance)
#    specific statement, fits owner 0.10 / others 0.02: accept 0.651 vs 0.183
#    specific statement, fits owner 0.30 / others 0.02: accept 0.972 vs 0.183

f=0.10f = 0.10 的陈述,对照 1 个经历时只有 0.100 的机会被接受,对照 10 个经历时为 0.651,对照 30 个时为 0.958;f=0.20f = 0.20 时对照 10 个经历已达 0.893。人的记忆里可供对照的经历远不止 30 个。一份由 10 句 f=0.10f = 0.10 的陈述组成的解读,在 M=10M = 10 时平均被接受 6.51 句;而把这份解读交给任何一个别人,被接受的句数完全相同。因此,若请当事人从若干份解读中认出自己的那一份,笼统陈述只能给出机会水平:两份中挑一为 0.5000,三份中挑一为 0.3333,四份中挑一为 0.2500。

只有具体的陈述才有区分力:若一句话对当事人的符合比例为 0.10、对他人为 0.02,接受率就是 0.651 对 0.183;这才是一个可以检验的差别。

覆盖模型:陈述的笼统程度 f、对照的经历数 M 与被接受的概率(p13)。

选择性记忆与事后改释

第二个机制发生在解读之后。设每句陈述碰巧为真的概率为 hh;命中被记住的概率为 a=0.9a = 0.9,落空被记住的概率为 b=0.4b = 0.4;落空的陈述还有 qq 的机会被事后改释为「以另一种方式应验」。记忆中的准确率为

ha+(1−h)bqha+(1−h)b.\frac{h a + (1-h) b q}{h a + (1-h) b}.

推导

# Chapter 6. p14: Remembered accuracy - selective recall and retrofitting.
# Each of 10 statements in a reading is true by chance with probability h.
# Afterwards a hit is remembered with probability a, a miss with
# probability b < a, and a miss may be reinterpreted as a hit ("it came
# true in another way") with probability q. Remembered accuracy:
#   (h a + (1 - h) b q) / (h a + (1 - h) b).
# A simulation with a fixed seed checks the formula.
import random

def remembered(h, a, b, q):
    hit = h * a + (1 - h) * b * q
    return hit / (h * a + (1 - h) * b)

if __name__ == "__main__":
    a, b = 0.9, 0.4
    print("   chance hit rate h   q = 0.0   q = 0.2   q = 0.4")
    for h in (0.2, 0.3, 0.5):
        print("   {:17.2f}   {:7.3f}   {:7.3f}   {:7.3f}".format(
            h, *(remembered(h, a, b, q) for q in (0.0, 0.2, 0.4))))
    rng = random.Random(71014)
    h, q, T = 0.3, 0.2, 100000
    num = den = 0
    for _ in range(T):
        for _ in range(10):
            true = rng.random() < h
            if rng.random() < (a if true else b):
                den += 1
                num += true or rng.random() < q
    print("   simulation, h = 0.30, q = 0.20: {:.3f} (formula {:.3f})".format(
        num / den, remembered(h, a, b, q)))
    print("   readings with 5+ chance hits of 10 at h = 0.30: {:.4f}".format(
        sum(__import__("math").comb(10, k) * h ** k * (1 - h) ** (10 - k)
            for k in range(5, 11))))
#    chance hit rate h   q = 0.0   q = 0.2   q = 0.4
#                 0.20     0.360     0.488     0.616
#                 0.30     0.491     0.593     0.695
#                 0.50     0.692     0.754     0.815
#    simulation, h = 0.30, q = 0.20: 0.592 (formula 0.593)
#    readings with 5+ chance hits of 10 at h = 0.30: 0.1503

碰巧命中率为 0.30 时,仅选择性记忆就把记住的准确率抬到 0.491;再加上 0.2 的改释概率,升到 0.593(模拟值 0.592)。碰巧命中率为 0.50 时,两者分别为 0.692 与 0.754。另外,即使每句只有 0.30 的机会碰巧为真,十句中有五句以上「说中」的解读也占 0.1503,这些解读最容易被记住和转述。

假说 · 尚未验证

作者的假说(尚未验证): 当事人对一次解读的事后准确度评分,主要由陈述的笼统程度与选择性记忆决定,而不由牌面决定。什么会证明它错: 若在第八章的盲法配对检验中,当事人认出自己解读的比例显著高于机会水平,而且在只保留具体陈述的版本中依然如此,这一假说就不成立。

小结

笼统陈述在丰富的记忆面前几乎必然被接受,而且对任何人都同样被接受,因此不能区分解读的对象;选择性记忆与事后改释再把记住的准确率抬高一截。检验解读的正确方式不是询问「准不准」,而是盲法配对。

参考文献

Dickson, D. H., and I. W. Kelly. 1985. “The ‘Barnum Effect’ in Personality Assessment: A Review of the Literature.” Psychological Reports 57 (2): 367–82.
Forer, Bertram R. 1949. “The Fallacy of Personal Validation: A Classroom Demonstration of Gullibility.” The Journal of Abnormal and Social Psychology 44 (1): 118–23.

Fan, Y. (2026). Tarot Hypergeometric Sampling, Lenormand Grid Topology and Kabbalistic Tree Graphs: Reversal Binomial Mixtures, Grand-Tableau Manhattan Distances and Twenty-Two-Edge Spectral Centrality [塔罗超几何抽样、雷诺曼网格拓扑与卡巴拉生命树图论] (Version 1.0.0). 5D Library. https://doi.org/10.5281/zenodo.24090191 · Licensed under Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International (CC BY-NC-ND 4.0).