5D Library › 回溯信念锁定、序贯痛点唤醒与马尔可夫升单自动机 · English

第一章 回溯断言与「信任雪崩」

一条命中的断言值多少

测算服务中最有说服力的时刻,往往不是预测未来,而是说中过去:「2020 年前后你经历过一次很大的低谷」「你小时候和父亲的关系很紧张」。对一位怀疑的用户来说,这样的「回溯」比任何预言都容易检验,因此也更容易建立信任。

贝叶斯定理给出这种信任应当增长多少 (Good 1950)。设用户起初认为测算师具有真实洞察力的概率为 0.01(赔率 1 比 99)。一条断言若测算师确有洞察力,命中的概率为 hh;若没有,命中的概率等于它在人群中的基率 bb。一次命中把赔率乘以似然比 LR=hb,log⁡oddsm=log⁡odds0+∑i=1mlog⁡LRi.\mathrm{LR} = \frac{h}{b} ,\qquad \log \mathrm{odds}_m = \log \mathrm{odds}_0 + \sum_{i=1}^{m} \log \mathrm{LR}_i . 信任是否「雪崩」,完全取决于 bb:同样是命中,基率高的断言几乎不提供证据。

证据

巴纳姆效应。 Forer 让学生做一份人格测验,然后发给每人同一段从星座书中摘来的泛泛描述,学生平均把它评为对自己非常准确 (Forer 1949)。这说明人们会把基率很高的断言当作专门针对自己的断言,也就是把 bb 判断得过低。

推导

# Chapter 1. [p01] What a retrodictive "hit" is worth: likelihood ratios.
# A sceptic gives the reader a prior probability of real insight of
# 0.01 (odds 1:99). A retrodictive statement about her past hits with
# probability h if the reader has insight and with the base rate b if
# not; a hit multiplies the odds by LR = h / b. Four kinds of
# statement. A generic statement feels personal (the Forer effect):
# the user takes its base rate to be 0.20, not 0.80. For the
# "trough year" the user believes b = 1/15 (one of
# 15 years) while the true base rate is 0.25, because in the population
# one year (a shared shock) is the worst year for a quarter of people.
# Hits needed to reach a posterior of 0.5 and of 0.95, with true and
# with perceived base rates.
import math

PRIOR = 0.01
KINDS = (("generic (Barnum)", 0.95, 0.80, 0.20),
         ("semi-specific", 0.90, 0.30, 0.30),
         ("trough year of 15", 0.90, 0.25, 1 / 15),
         ("checkable detail", 0.90, 0.02, 0.02))

def hits_needed(lr, target):
    need = math.log(target / (1 - target)) - math.log(PRIOR / (1 - PRIOR))
    return math.ceil(need / math.log(lr)) if lr > 1 else None

if __name__ == "__main__":
    print("   statement             h     b true  b seen  LR true  LR seen")
    for name, h, b, seen in KINDS:
        print("   {:19s} {:5.2f} {:8.3f} {:7.3f} {:8.2f} {:8.2f}".format(
            name, h, b, seen, h / b, h / seen))
    print("\n   hits needed for posterior   0.5 (true/seen)"
          "  0.95 (true/seen)")
    for name, h, b, seen in KINDS:
        print("   {:19s} {:>12s} {:>17s}".format(
            name,
            "{}/{}".format(hits_needed(h / b, 0.5),
                           hits_needed(h / seen, 0.5)),
            "{}/{}".format(hits_needed(h / b, 0.95),
                           hits_needed(h / seen, 0.95))))
    print("\nReading: a generic statement is almost no evidence (LR 1.19);")
    print("a sceptic would need 27 such hits, but if it feels personal she")
    print("is half convinced after 3 and sure after 5. The trough year")
    print("looks strong for the same reason. Only checkable details carry")
    print("the weight they seem to.")
#    statement             h     b true  b seen  LR true  LR seen
#    generic (Barnum)     0.95    0.800   0.200     1.19     4.75
#    semi-specific        0.90    0.300   0.300     3.00     3.00
#    trough year of 15    0.90    0.250   0.067     3.60    13.50
#    checkable detail     0.90    0.020   0.020    45.00    45.00
#
#    hits needed for posterior   0.5 (true/seen)  0.95 (true/seen)
#    generic (Barnum)            27/3              44/5
#    semi-specific                5/5               7/7
#    trough year of 15            4/2               6/3
#    checkable detail             2/2               2/2
#
# Reading: a generic statement is almost no evidence (LR 1.19);
# a sceptic would need 27 such hits, but if it feels personal she
# is half convinced after 3 and sure after 5. The trough year
# looks strong for the same reason. Only checkable details carry
# the weight they seem to.

泛泛的断言(基率 0.80)的真实似然比只有 1.19,一位先验为 0.01 的怀疑者需要 27 次命中才会达到半信,44 次才会确信;但只要它「感觉很准」,她把基率当成 0.20,似然比就变成 4.75,3 次命中就半信,5 次就确信。「过去 15 年里的波谷年份」看起来是十五选一,似然比 13.50;可是如果某一年(例如一场共同的冲击)是四分之一人群的最差年份,真实基率是 0.25,似然比只有 3.60。真正有分量的只有可以核对的细节:基率 0.02,似然比 45.00,两次命中就足以让怀疑者确信。

「雪崩」是怎样发生的

程序模拟 20,000 位怀疑的用户,先验对数赔率围绕 0.01 上下分散。测算师没有任何洞察力,只说泛泛的断言,每条以基率 0.80 命中。一组用户按真实似然比更新,另一组按感受到的似然比更新,并且几乎忘记说错的断言(只记五分之一的分量)。

推导

# Chapter 1. [p02] A "trust avalanche" without any insight.
# 20,000 sceptical users (seed 69); prior log-odds of insight normal
# with mean log(1/99) and sd 1.5. A reader WITHOUT insight makes m
# generic retrodictive statements; each hits with its base rate 0.80.
# Proper updating uses the true likelihood ratios (hit 0.95/0.80, miss
# 0.05/0.20). Biased updating uses the perceived base rate 0.20 (hit
# 0.95/0.20) and gives each miss only weight 0.2 (misses are
# forgotten). Reported: share of users whose posterior exceeds 0.9.
import math
import random

N = 20000

def run(m, biased, rng):
    lh = math.log(0.95 / (0.20 if biased else 0.80))
    lm = math.log(0.05 / (0.80 if biased else 0.20))
    wm = 0.2 if biased else 1.0
    sure = 0
    for _ in range(N):
        lo = rng.gauss(math.log(1 / 99), 1.5)
        for _ in range(m):
            lo += lh if rng.random() < 0.80 else wm * lm
        sure += lo > math.log(9)
    return sure / N

if __name__ == "__main__":
    print("    m   share sure (proper)   share sure (biased)")
    for m in (0, 1, 2, 3, 4, 5, 6, 8, 10, 15):
        a = run(m, False, random.Random(69))
        b = run(m, True, random.Random(69))
        print("   {:2d} {:20.4f} {:21.4f}".format(m, a, b))
    print("\nReading: with the true likelihood ratios almost nobody is")
    print("convinced by any number of generic hits. With the felt ratios")
    print("the share jumps from a few per cent to most users within a")
    print("handful of statements: an avalanche made of misjudged base")
    print("rates, not of evidence.")
#     m   share sure (proper)   share sure (biased)
#     0               0.0000                0.0000
#     1               0.0000                0.0003
#     2               0.0000                0.0043
#     3               0.0000                0.0408
#     4               0.0000                0.1603
#     5               0.0001                0.3452
#     6               0.0000                0.5263
#     8               0.0000                0.7928
#    10               0.0001                0.9261
#    15               0.0001                0.9960
#
# Reading: with the true likelihood ratios almost nobody is
# convinced by any number of generic hits. With the felt ratios
# the share jumps from a few per cent to most users within a
# handful of statements: an avalanche made of misjudged base
# rates, not of evidence.
给出的泛泛回溯断言数与确信(后验大于 0.9)的用户比例。

按真实似然比更新的用户,无论听多少条泛泛断言,确信的比例都几乎为零。按感受更新的用户却呈现典型的 S 形:3 条时 0.0408,6 条时 0.5263,10 条时 0.9261。这正是提纲所说的「相变坍缩」,但它是由误判的基率与遗忘的失误构成的,不是由证据构成的。

假说 · 尚未验证

作者的诠释(尚未验证): 测算服务中用户的「被说中」体验,大部分可以由泛泛断言、基率误判与选择性回忆解释,与测算方法是否有效无关。什么会证明这一判断错:若在第八章那样的盲法配对检验中,用户能以显著高于三分之一的比例认出为自己写的回溯断言,说明断言中包含了泛泛断言之外的信息。

小结

一条回溯断言值多少,取决于它的基率。泛泛的断言几乎不是证据,但会被感受为证据;「波谷年份」若是人群共同的低谷,也远不如看上去特异。所谓信任雪崩,在本章的模型中完全来自误判的基率与遗忘的失误。

参考文献

Forer, Bertram R. 1949. “The Fallacy of Personal Validation: A Classroom Demonstration of Gullibility.” Journal of Abnormal and Social Psychology 44 (1): 118–23.
Good, I. J. 1950. Probability and the Weighing of Evidence. Charles Griffin.

Fan, Y. (2026). Retrodictive Belief Locking, Sequential Pain-Point Agitation and Markovian Upsell Automata: Bayesian Trust Avalanches, Sunk-Cost Commitment Ratchets and Cross-Domain Transition Matrices [回溯信念锁定、序贯痛点唤醒与马尔可夫升单自动机] (Version 1.0.0). 5D Library. https://doi.org/10.5281/zenodo.24090186 · Licensed under Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International (CC BY-NC-ND 4.0).