关于「增量实验:留出组、地区实验与沙普利分配」,《暗投、归因与双轨抽佣》给出的定量结论是什么?
结论:留出组所需人数随人均广告费的平方反比增长:变异系数 10、人均 RM 0.02 时,仅分辨增量 ROAS 是否达到盈亏线,每组就要 1,061,621 人。幽灵广告式的设计能省掉曝光稀释带来的 11.1 倍,但需要平台支持。全马十六个地区的地区实验在预算比例 0.03 时功效只有 0.734,比例加倍到 0.06 时升到 0.999;一次实验估计出 2.390,精确随机化检验的 p 值是 0.0005。析因实验的沙普利值把 RM 288,000 的增量分成 124,500、102,000 与 61,500,合计不多不少;但 A 在其余渠道都开着时的回报是 3.600,高于沙普利值的 3.112,分钱与花钱要用两把尺子。
数据口径:在一个所有真相都已知的合成周里,三个后台认领的订单合计是实际订单的 1.112 倍,真正由广告带来的增量只占实际订单的 0.312;后台 ROAS 之和是 4.81,真实增量 ROAS 只有 1.35,低于盈亏平衡线 2.884。 核对原文之后,本书做了蓝图修正:付费向自然流量的外溢在随机实验里有正有负,本书把它写成可正可负的净效应;沙普利值适合分功劳,不适合定预算;用户级留出组对多数东南亚品牌太大,重点改为全马十六个地区的地区实验;平台费率无法逐项核实,全部作为示意输入;马来西亚电子交易条例已由 2024 年新条例取代。
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: 0.0005, 0.03, 0.06, 0.734, 0.999, 2.390
spend share 0.03, true iROAS 0.000: mean -0.010, sd 1.252, one-sided power 0.050 spend share 0.03, true iROAS 2.884: mean +2.825, sd 1.265, one-sided power 0.734 spend share 0.06, true iROAS 0.000: mean +0.016, sd 0.625, one-sided power 0.050 spend share 0.06, true iROAS 2.884: mean +2.889, sd 0.655, one-sided power 0.999 one experiment at share 0.06: estimate 2.390; exact randomisation p over 12,870 assignments 0.0005
核心公式
查看完整验算程序
# Chapter 5. p52: a geo experiment over Malaysia's 16 states and federal territories (synthetic).
# Weekly GMV of region i is BASE * w_i * season_t * (1 + trend_i * t) * noise. Eight regions get
# whitelisting spend of a share k of their baseline GMV for 4 weeks after 4 pre-weeks. Estimator:
# incremental GMV = sum over treated of (post - pre * control growth), divided by treated spend.
# Region weights are rough illustrative shares, not census figures.
import itertools
import math
import random
rng = random.Random(16)
W = [0.20, 0.12, 0.10, 0.08, 0.08, 0.07, 0.06, 0.06, 0.06, 0.05, 0.04, 0.04, 0.03, 0.006, 0.004, 0.004]
BASE, PRE, POST = 400000.0, 4, 4
SEASON = [1.00, 0.97, 1.05, 1.10, 1.02, 0.99, 1.08, 1.12]
def world(r, treated, k):
pre, post, spend = [], [], []
for i, w in enumerate(W):
tr = rng.gauss(0.0, 0.01)
g = [BASE * w * SEASON[t] * (1.0 + tr * t) * math.exp(rng.gauss(0.0, 0.06)) for t in range(8)]
s = k * BASE * w * POST if i in treated else 0.0
pre.append(math.fsum(g[:PRE]))
post.append(math.fsum(g[PRE:]) + r * s)
spend.append(s)
return pre, post, spend
def estimate(pre, post, treated, k):
ctrl = [i for i in range(16) if i not in treated]
growth = math.fsum(post[i] for i in ctrl) / math.fsum(pre[i] for i in ctrl)
inc = math.fsum(post[i] - pre[i] * growth for i in treated)
return inc / math.fsum(k * BASE * W[i] * POST for i in treated)
for k in (0.03, 0.06):
for r in (0.0, 2.884):
est = []
for _ in range(1000):
t = set(rng.sample(range(16), 8))
pre, post, _ = world(r, t, k)
est.append(estimate(pre, post, t, k))
mean = math.fsum(est) / len(est)
sd = math.sqrt(math.fsum((x - mean) ** 2 for x in est) / (len(est) - 1))
if r == 0.0:
crit = sorted(est)[949]
power = sum(1 for x in est if x > crit) / len(est)
print("spend share {:.2f}, true iROAS {:.3f}: mean {:+.3f}, sd {:.3f}, one-sided power {:.3f}".format(
k, r, mean, sd, power))
t = set(rng.sample(range(16), 8))
pre, post, _ = world(2.884, t, 0.06)
obs = estimate(pre, post, t, 0.06)
null = [estimate(pre, post, set(c), 0.06) for c in itertools.combinations(range(16), 8)]
p = sum(1 for x in null if x >= obs) / len(null)
print("one experiment at share 0.06: estimate {:.3f}; exact randomisation p over {:,} assignments {:.4f}"
.format(
obs, len(null), p))唯一数据源与引用
范雅福(Fan, Y.). (2026). 《东南亚社交电商暗投带货联动、站内外归因叠加与双轨抽佣增长母典》第 5 章「增量实验:留出组、地区实验与沙普利分配」(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-351 · v1.0.0 · CC-BY-NC-ND-4.0
BibTeX
@book{fan2026351,
author = {Fan, Yafu},
title = {Southeast Asian Social-Commerce Whitelist Amplification, Cross-Channel Attribution and Dual-Track Commission Growth},
year = {2026},
publisher = {5D Library},
edition = {1.0.0},
note = {DOI pending}
}