所属模块：
M1 · 数学与统计基础 (Mathematics & Statistics Fundamentals)| 专题分类：实验设计 (A/B) (实验设计 (A/B))| 难度等级：Hard
一、核心一句话结论 (One-Sentence Summary)
只分析’实际受处理影响’的子群（被触发器激活的用户），提升功效并聚焦机制。
Trigger analysis evaluates treatment effects exclusively among the sub-population of users who actually encountered the experimental trigger condition in both groups, filtering out un-exposed noise and dramatically boosting statistical power.
二、核心考点要义 (Key Insights)
- 📌 适合处理只影响部分用户的场景（如只影响特定页面）
- 📌 需在实验前定义触发器且不能依赖处理
English Insights:
– Core Intuition: If a new feature is located deep inside the checkout settings page, 95% of users never visit that page. Pooling all users dilutes the true effect by factor 20.
– Dual Triggering: Must identify users in BOTH Treatment and Control who satisfied the trigger condition (e.g. Navigated to settings page), preserving randomized balance.
– Mathematical Consistency: Under the assumption of zero effect on untriggered users, Triggered Effect = $frac{text{Overall Effect}}{P(text{Triggered})}$, with standard error reduced by factor $approx sqrt{P(text{Triggered})}$.
三、核心数学原理与机理推导 (Mathematical Principles & Derivation)
$$text{trigger}: text{subset actually exposed to the treatment}$$
问题的背景：很多处理只影响部分用户——例如’优化搜索框的排序’只影响发起搜索的用户，’改进结账页’只影响进入结账流程的用户。若对全体用户分析，大部分用户的指标不受影响（他们的指标只是噪声），导致效应被稀释（观测到的平均效应 = 真实效应 × 受影响比例），功效大幅下降。触发器分析（trigger analysis / exposed analysis） 的做法是：只在’实际受处理影响’的子群（触发器激活的用户）上分析。统计上的关键区别：触发器必须是实验前可定义的、且不受处理影响的变量（如’是否发起搜索’——用户的搜索意图与处理无关）。此时触发组内的分析是无偏的（因为触发器与处理分配独立），且功效更高（效应不被稀释）。反之，若用处理后的变量定义子群（如’是否点击了按钮’），则会产生选择偏差（处理改变了谁进入子群），这类分析被称为’条件化在结果上’，是严重错误。
📖 查看英文严格数学推导 (English Mathematical Derivation)
Let $E_i in {0, 1}$ indicate whether user $i$ triggered the feature surface. By the law of total expectation, overall population treatment effect is: $tau_{text{all}} = E[Y(1) – Y(0)] = P(E=1) E[Y(1) – Y(0) mid E=1] + P(E=0) E[Y(1) – Y(0) mid E=0]$. Under the exclusion restriction that untriggered users experience zero effect ($E[Y(1) – Y(0) mid E=0] = 0$), we have $tau_{text{all}} = p_{text{trig}} cdot tau_{text{trig}}$. The variance of the naive overall estimator is $text{Var}(hat{tau}_{text{all}}) approx frac{2sigma^2}{N}$. If we evaluate only triggered users ($N_{text{trig}} = N p_{text{trig}}$), the variance of $hat{tau}_{text{trig}}$ is $frac{2sigma^2}{N p_{text{trig}}}$. Transforming back to population scale: $text{Var}(p_{text{trig}} hat{tau}_{text{trig}}) = p_{text{trig}}^2 left(frac{2sigma^2}{N p_{text{trig}}}right) = p_{text{trig}} left(frac{2sigma^2}{N}right) = p_{text{trig}} text{Var}(hat{tau}_{text{all}})$, reducing variance by exactly factor $p_{text{trig}}$.
四、工业级落地权衡与工程考量 (Industrial Trade-offs)
实践要点：① 触发器 vs 后分层——触发器是实验前定义的（无偏），后分层是实验后按协变量加权（也需实验前协变量）；两者的共同要求是不能用处理后的变量。② 触发器 vs CUPED——触发器提升功效的方式是’剔除不受影响的样本’（提高信噪比），CUPED 是’用协变量降方差’；两者可叠加。③ 触发器的常见类型——(a) 页面/功能触发（访问了被改动的页面）；(b) 意图触发（发起搜索、打开结账）；(c) 设备触发（移动端用户，若改动只影响移动端）。④ 报告规范——应同时报告’全体效应’与’触发组效应’：全体效应反映业务总体影响（决策依据），触发组效应反映机制效应（理解为何有效）；只报触发组效应会夸大业务影响。⑤ 稀释比例的估计——触发组占比可从数据估计，用于把触发组效应换算为全体效应（全体效应 ≈ 触发组效应 × 触发比例）。⑥ 陷阱——(a) 用处理后变量定义触发器（选择偏差）；(b) 只报触发组效应而误导业务（夸大）；(c) 触发器定义过于宽泛（稀释仍严重）或过窄（样本不足）。
⚙️ 查看英文落地权衡分析 (English Systems & Trade-offs)
Triggering implementation protocol: Logging the trigger event must be executed counterfactually in the control group. For example, if Treatment adds a modal popup upon clicking a button, Control must log a hidden event ‘would_have_triggered_modal’ at that exact button click without showing the popup. If triggering is only logged when the Treatment UI renders, Control has no trigger events, making direct comparison impossible without catastrophic selection bias.
五、常见面试避坑陷阱 (Common Pitfalls & Traps)
- ⚠️ 用处理后的行为（如是否点击）定义触发器（选择偏差）
- ⚠️ 只报触发组效应而不报全体效应
English Pitfalls:
– Conditioning on a post-treatment trigger that is itself influenced by treatment (e.g. A trigger that Treatment makes easier to reach), which introduces severe selection bias and invalidates randomization.
– Failing to implement virtual / counterfactual trigger logging in the Control group.
六、高频深度面试追问与预测 (Follow-Up Questions)
- 触发器分析与 CUPED 的区别？
- Why does counterfactual trigger logging in Control preserve exact internal validity?
- 如何避免用结果定义触发器？
- How does Trigger Analysis relate to Complier Average Causal Effect (CACE) in Instrumental Variable analysis?
七、知识图谱对齐 (Knowledge Graph Anchor)
- 🔗 关联底层卡片：
工业级 A/B 实验设计、分流正交、SRM 卡方排查与方差缩减(Industrial A/B Testing: Split, SRM & Variance Reduction) - 🗺️ 知识图谱模块：
数据科学与因果实验导图
🔬 算法科学家与机器学习深度考察全量题库 (Science Depth)
本题收录于 TalentMe 算法科学家深度考察真题库 (Science Depth)。全库共 856 道硬核考点，深度覆盖数学统计、经典ML、深度学习、Transformer、大语言模型、多模态、推荐系统与 MLOps。支持 Jev 面经智能匹配、一键离线单文件 HTML 手册导出并直连 Obsidian 本地记忆。