5D Library › 超临界病毒级联、信息缺口风险率与多层付费墙优化 · English

第八章 四步核查、合规清单与一项预注册的付费墙检验

相信一个增长计划之前的四个问题

本章把全书变成四个问题,可以向任何裂变或付费墙方案——包括作者自己的——提出。它们不是经营建议。

一、K 是怎么测的? 按树状结构估算的 K 在真实网络中偏高(第二章),按人数计算的 K 会被假账号抬高(第三章)。要看按推荐日志逐代测得的 K、去除假账号后的 K,以及单个种子的熄灭比例(第一章)。

二、获客成本算全了吗? 免费测算的算力与送达成本、推荐分佣、对假账号的稽核,都是获客成本(第二、三章)。要看每个付费用户的完整获客成本,而不是每个注册用户的广告费。

三、多收的钱从哪里来? 对一件商品,菜单胜不过单一价格(第五章)。多层付费墙多收的钱应当能指认为多卖的商品或捆绑;诱饵档的效果要单独测量(第六章)。

四、转化率衡量的是什么? 转发率与「说得准」的评分衡量的是感觉(第四章)。要用盲法调包检验衡量解读是否区分了人。

合规清单(需按经营所在地核对,不是法律意见)

  1. 分享后解锁:一些平台把「分享后才能查看结果」列为禁止的诱导分享,并限制测试类外链 (ifanr, n.d.)。
  2. 分佣层级:只做单层推荐奖励;按下线的下线计酬可能构成多级计酬 (State Council of the People’s Republic of China 2005)。
  3. 广告内容:在中国大陆,《广告法》第九条禁止广告含有迷信内容 (Standing Committee of the National People’s Congress 2015);其他地区对算命类广告也可能有特别规定。
  4. 个性化定价:用个人数据对同一商品向不同的人报不同价格,可能违反关于自动化决策的规定 (Standing Committee of the National People’s Congress 2021);公开菜单的自选档位风险较低。
  5. 反复收费与焦虑营销:利用焦虑或好奇心衰减促使用户为同一件事反复付费,可能违反消费者保护规定。

有推荐时怎样做实验

比较两种付费墙时,按访客随机分组会出问题:被邀请的朋友看到的版本取决于邀请者,两组互相污染。补救办法是按整棵推荐树随机,代价是设计效应 1+(m‾−1)ρ1 + (\bar m – 1)\rho,其中 m‾\bar m 是推荐树的平均规模,ρ\rho 是树内相关系数。

推导

# Chapter 8. Testing a paywall when users recruit each other.
# Comparing two paywall designs by randomising individual visitors fails
# when users invite friends: an invited friend's behaviour depends on the
# inviter's version (interference). The remedy is to randomise whole
# referral trees - each seed and everyone it brings in gets the seed's
# version - and to pay for it with a design effect
#   DE = 1 + (mean tree size - 1) x ICC,
# ICC the intra-tree correlation of the outcome. Outcome here: whether a
# visitor pays within 30 days, 5.0 % under A; smallest effect of interest
# +1.0 point (6.0 % under B); two-sided alpha 0.05, power 80 %.
# Part 2 shows the bias of individual randomisation when a share c of
# visitors in arm A arrived through a B-arm inviter and behave like B.
import math
from statistics import NormalDist

Z = NormalDist()

def n_individual(p1, p2, alpha=0.05, power=0.8):
    za, zb = Z.inv_cdf(1 - alpha / 2), Z.inv_cdf(power)
    pbar = (p1 + p2) / 2
    num = (za * math.sqrt(2 * pbar * (1 - pbar))
           + zb * math.sqrt(p1 * (1 - p1) + p2 * (1 - p2))) ** 2
    return math.ceil(num / (p2 - p1) ** 2)

if __name__ == "__main__":
    n = n_individual(0.05, 0.06)
    print("   individual randomisation: {:,} visitors per arm".format(n))
    print("\n   mean tree size   ICC    design effect   visitors per arm")
    for size in (3, 8):
        for icc in (0.02, 0.05, 0.20):
            de = 1 + (size - 1) * icc
            print("        {:2d}         {:4.2f}       {:5.2f}          {:,}"
                  .format(size, icc, de, math.ceil(n * de)))
    print("\n   contaminated share c   measured difference (true +1.0)")
    for c in (0.0, 0.1, 0.3, 0.5):
        print("          {:3.1f}                  {:+.2f} points".format(
            c, 1.0 * (1 - c)))
    print("\nReading: randomising referral trees costs a design effect that")
    print("can more than double the sample; randomising individuals instead")
    print("shrinks the measured effect towards zero by the share of")
    print("visitors who arrived through the other arm.")
#    individual randomisation: 8,158 visitors per arm
#
#    mean tree size   ICC    design effect   visitors per arm
#          3         0.02        1.04          8,485
#          3         0.05        1.10          8,974
#          3         0.20        1.40          11,422
#          8         0.02        1.14          9,301
#          8         0.05        1.35          11,014
#          8         0.20        2.40          19,580
#
#    contaminated share c   measured difference (true +1.0)
#           0.0                  +1.00 points
#           0.1                  +0.90 points
#           0.3                  +0.70 points
#           0.5                  +0.50 points
#
# Reading: randomising referral trees costs a design effect that
# can more than double the sample; randomising individuals instead
# shrinks the measured effect towards zero by the share of
# visitors who arrived through the other arm.

按个人随机时,要以 80% 的把握看见付费率从 5.0% 升到 6.0%,每组需要 8,158 名访客。按推荐树随机时,平均树规模为 8、树内相关为 0.20 时,设计效应为 2.40,每组需要 19,580 名访客。如果仍按个人随机,一半访客经由另一组的邀请者到来时,测得的差异只剩真实差异的一半(+0.50 个百分点)。

一项预注册的付费墙检验

假说 · 尚未验证

研究方案。 比较 A(完整测算单一价格 9.9)与 B(分层菜单)。随机单位为推荐树。主要结局:每个种子 30 天内的净收入,计入树内全部访客:付款减去分佣,减去送出的免费测算成本(每份 0.15)。主要估计量 DD:B 组均值减 A 组均值;最小关注效应为每个种子 0.10。三道闸门使结果不可解释而非否定:推荐树分组比例失衡(|z|>3|z| > 3)、超过 5% 的推荐树混入了两个版本、入组不足计划的 90%。由独立于定价决策的人员按登记的代码实施与分析,结果不论正负都发表。

推导

# Chapter 8. A pre-registered paywall test and its decision rule.
# Arms: A = one price for the full reading (9.9); B = the tiered menu.
# Unit of randomisation: the referral tree (p15). Primary outcome: NET
# revenue per seed visitor over 30 days, counting every visitor in the
# seed's tree: payments minus commissions paid out minus the cost of the
# free readings given (0.15 each). Estimand D = mean(B) - mean(A); the
# smallest effect of interest is 0.10 per seed. Gates that make a result
# uninterpretable rather than negative: sample-ratio mismatch (|z| > 3 on
# the split of trees), more than 5 % of trees exposed to both arms, and
# fewer than 90 % of the planned trees. The program sizes the test by
# simulating trees (outcome per seed: a heavy-tailed sum, as in p02; a
# tenth of payments paid out as commission; B lifts paying by 0.4 point)
# and prints the decision for each branch of the rule.
import math
import random
from statistics import NormalDist

Z = NormalDist()

def tree_outcome(rng, uplift):
    size = 1
    alive = 1
    while alive and size < 400:
        kids = sum(rng.random() < 0.18 for _ in range(5))
        alive += kids - 1
        size += kids
    pays = sum(rng.random() < 0.05 + uplift for _ in range(size))
    return pays * 9.9 * 0.9 - 0.15 * size

def decide(d, se, srm_z, mixed, enrolled):
    if abs(srm_z) > 3:
        return "uninterpretable: sample-ratio mismatch"
    if mixed > 0.05:
        return "uninterpretable: arms mixed in more than 5 % of trees"
    if enrolled < 0.9:
        return "uninterpretable: under 90 % of planned trees"
    lo, hi = d - 1.96 * se, d + 1.96 * se
    if lo > 0.10:
        return "B better by at least the smallest effect of interest"
    if hi < 0:
        return "B worse: the tiered menu earns less per seed"
    if lo > 0:
        return "B better, but not shown to reach 0.10"
    return "no difference shown"

if __name__ == "__main__":
    rng = random.Random(63016)
    a = [tree_outcome(rng, 0.0) for _ in range(20000)]
    b = [tree_outcome(rng, 0.004) for _ in range(20000)]
    ma, mb = sum(a) / len(a), sum(b) / len(b)
    sd = math.sqrt((sum((x - ma) ** 2 for x in a)
                    + sum((x - mb) ** 2 for x in b)) / (len(a) + len(b) - 2))
    n = math.ceil(2 * ((1.96 + Z.inv_cdf(0.8)) * sd / 0.10) ** 2)
    print("   simulated net revenue per seed: A {:.3f}, B {:.3f}".format(
        ma, mb))
    print("   pooled SD {:.3f}  ->  {:,} trees per arm for 80 % power"
          .format(sd, n))
    se = sd * math.sqrt(2 / n)
    print("\n   decision rule, one line per branch:")
    for case in ((0.25, 0.0, 0.0, 1.0), (0.12, 0.0, 0.0, 1.0),
                 (0.05, 0.0, 0.0, 1.0), (-0.20, 0.0, 0.0, 1.0),
                 (0.25, 3.5, 0.0, 1.0), (0.25, 0.0, 0.08, 1.0),
                 (0.25, 0.0, 0.0, 0.8)):
        print("   D = {:+.2f}: {}".format(case[0], decide(case[0], se,
                                                       *case[1:])))
    print("\nReading: because one seed's tree may hold hundreds of users,")
    print("the outcome is so dispersed that detecting 0.10 per seed takes")
    print("well over a hundred thousand trees per arm. A test that small")
    print("sites cannot afford should not be replaced by a guess.")
#    simulated net revenue per seed: A 2.987, B 3.253
#    pooled SD 10.308  ->  166,788 trees per arm for 80 % power
#
#    decision rule, one line per branch:
#    D = +0.25: B better by at least the smallest effect of interest
#    D = +0.12: B better, but not shown to reach 0.10
#    D = +0.05: no difference shown
#    D = -0.20: B worse: the tiered menu earns less per seed
#    D = +0.25: uninterpretable: sample-ratio mismatch
#    D = +0.25: uninterpretable: arms mixed in more than 5 % of trees
#    D = +0.25: uninterpretable: under 90 % of planned trees
#
# Reading: because one seed's tree may hold hundreds of users,
# the outcome is so dispersed that detecting 0.10 per seed takes
# well over a hundred thousand trees per arm. A test that small
# sites cannot afford should not be replaced by a guess.

模拟中两组每个种子的净收入为 2.987 与 3.253,合并标准差为 10.308——一个种子的推荐树可能有几百人,结局极其分散。要以 80% 的把握看见 0.10 的差异,每组需要 166,788 棵推荐树。判定规则对每个分支打印一行:B 至少好出最小关注效应;B 更好但未证明达到 0.10;没有显示差异;B 更差;以及三种不可解释的情形。

注

结果意味着什么、不意味着什么。 它描述的是检验期间、在这两种设计下的推荐树。小网站做不起这样规模的检验时,不应当用猜测代替它;更好的做法是检验一个更大的效应,或者延长检验期并预先登记中期分析的规则。

小结

四个问题概括了全书:K 是否按推荐日志逐代测得并去除了假账号;获客成本是否计入免费测算、分佣与稽核;多收的钱是否来自多卖的商品;转化率与准确度评分衡量的是感觉还是区分力。合规清单列出了分享解锁、分佣层级、广告内容、个性化定价与反复收费五项需要核对的规则。有推荐时,实验应按推荐树随机;一项以每个种子净收入为主要结局的预注册检验,每组需要 166,788 棵推荐树。

参考文献

ifanr. n.d. Report on Enforcement of WeChat’s External Links Content Management Specification (Induced Sharing and H5 Test Content). In Chinese; News Report. https://www.ifanr.com/685476.
Standing Committee of the National People’s Congress. 2015. Advertising Law of the People’s Republic of China (Revised 2015), Article 9. In Chinese. https://www.wipo.int/wipolex/zh/text/370802.
Standing Committee of the National People’s Congress. 2021. Personal Information Protection Law of the People’s Republic of China, Article 24. In Chinese.
State Council of the People’s Republic of China. 2005. Regulations on Prohibiting Pyramid Selling (Order No. 444). In Chinese.

Fan, Y. (2026). Supercritical Viral Cascades, Information-Gap Hazard Rates and Multi-Tier Paywall Optimization: Galton-Watson Referral Thresholds, Share-to-Unlock Incentives and Second-Degree Price Discrimination [超临界病毒级联、信息缺口风险率与多层付费墙优化] (Version 1.0.0). 5D Library. https://doi.org/10.5281/zenodo.24090174 · Licensed under Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International (CC BY-NC-ND 4.0).