Applicable Scenarios
- New feature effect validation
- UI/UX redesign comparison
- Pricing strategy testing
- Recommendation algorithm comparison
- Landing page optimization
Core Principles
1. 先有假设再实验
"改变 X 会让 Y 提升 Z%"
2. 样本量要够
不够 = 噪音当信号
-
Run long enough At least 1~2 full business cycles
-
Check guardrail metrics Primary metric improves but others worsen = cannot launch
-
Statistical significance ≠ business significance p < 0.05 but effect 0.01% = meaningless
## 样本量计算
```text
公式因素:
- 基线转化率(p1)
- 最小检测效应(MDE)
- 显著性水平(α = 0.05)
- 统计功效(1-β = 0.8)
工具:
- Evan Miller 样本量计算器
- statsmodels(Python)
- power.prop.test(R)
Workflow
1. 明确假设和目标
2. 选择主指标 + 护栏指标
3. 计算样本量
4. 确定分流方式和时长
5. 启动实验
6. 等待(不提前偷看)
7. 分析结果
8. 做决策(全量 / 放弃 / 迭代)
Supporting Templates
templates/ab-test-template.md— A/B test design + results + conclusion
Quality Self-Check
□ 假设明确可证伪
□ 样本量计算
□ 分流无偏(hash)
□ 跑够时长
□ 不提前偷看
□ 看护栏指标
□ 结论可行动
Common Pitfalls
- Drawing conclusions with insufficient sample size——noise
- Peeking at p-value early——inflation
- Only looking at the primary metric——guardrail deterioration
- Insufficient duration——period effect
- Biased traffic splitting——uneven new users vs existing users
Collaboration with other skills
上游:metric-design → 指标选择
上游:product-manager → 假设来源
下游:analysis-report → 结果纳入报告