所属模块：
M1 · 数学与统计基础 (Mathematics & Statistics Fundamentals)| 专题分类：因果推断 (Causal Inference)| 难度等级：Easy
一、核心一句话结论 (One-Sentence Summary)
混淆需控制（否则有偏）；中介不应控制（会屏蔽效应）；对撞控制会引入选择偏差。
Confounders (forks $X leftarrow Z to Y$) cause spurious correlation and MUST be controlled; Mediators ($X to M to Y$) transmit mechanisms and should NOT be controlled for total effect; Colliders ($X to C leftarrow Y$) induce artificial association when controlled and MUST NOT be controlled.
二、核心考点要义 (Key Insights)
- 📌 用 DAG 判断该控制什么（后门准则）
- 📌 对撞偏差 = 条件化后制造出虚假相关（Berkson 悖论）
English Insights:
– Confounder ($Z$): Common cause of both treatment $X$ and outcome $Y$; creates non-causal backdoor paths that MUST be blocked/conditioned on.
– Mediator ($M$): Intermediate step along the causal pathway $X to M to Y$; controlling for $M$ blocks the indirect causal effect, leaving only direct effect.
– Collider ($C$): Common effect of both treatment $X$ and outcome $Y$; conditioning on a collider OPENS a spurious association path (Berkson’s Paradox).
三、核心数学原理与机理推导 (Mathematical Principles & Derivation)
$$text{confounder}: Xto T, Xto Y;quad text{collider}: Tto Xleftarrow Y$$
三类变量按 DAG 结构区分，控制策略完全相反：① 混淆变量（confounder）——同时影响处理 T 与结果 Y（X→T, X→Y），产生伪相关。必须控制，否则把 X 的效应误归给 T。后门准则给出系统判据：控制所有’从 T 到 Y 的后门路径’上的变量。② 中介变量（mediator）——位于因果链中间（T→M→Y），承载部分因果效应。不应控制：控制它会阻断该路径，得到的是’直接效应’而非’总效应’，从而低估总效应。若研究目的就是分解直接/间接效应，则需专门的中介分析而非简单控制。③ 对撞变量（collider）——同时被 T 与 Y 影响（T→C←Y）。不应控制：控制它会在 T 与 Y 之间制造虚假关联（对撞偏差 / collider bias / Berkson 悖论）。经典例子：录取 C 由天赋 X 与努力 Y 共同决定（X→C←Y），在录取者中控制 C 会得到’天赋与努力负相关’的虚假结论。
📖 查看英文严格数学推导 (English Mathematical Derivation)
DAG structural mechanics via d-separation: (1) Fork $X leftarrow Z to Y$: $X$ and $Y$ are dependent: $P(X, Y) ne P(X)P(Y)$. Conditioning on $Z$ blocks the path: $(X perp Y mid Z)$. Hence, adjusting for confounders isolates the true causal arc $X to Y$. (2) Chain $X to M to Y$: Conditioning on $M$ blocks the transmission: $(X perp Y mid M)$. Doing so eliminates the primary causal mechanism. (3) Collider $X to C leftarrow Y$: Unconditioned, $X$ and $Y$ are marginally independent: $(X perp Y)$. However, conditioning on $C$ induces correlation: $(X notperp Y mid C)$. For example, let $X$ = Talent, $Y$ = Hard Work, $C$ = College Admission ($C = X + Y$). Among admitted students ($C=1$), knowing someone is not talented implies they must be exceptionally hardworking, inducing artificial negative correlation.
四、工业级落地权衡与工程考量 (Industrial Trade-offs)
实践要点：① 先画 DAG 再决定控制什么——这是避免错误控制的最有效方法；仅凭’相关性’或’常识’选控制变量极易犯两类错误（漏控混淆、误控对撞）。② 中介分析的三种效应——总效应（不控制 M）、直接效应（控制 M）、间接效应（总−直接）；近年因果推断强调不要对中介做条件化，而应用 mediation analysis 或 path-specific effects。③ 对撞偏差的隐蔽性——它常在’看起来合理’的控制中发生，如按’是否住院’分层研究疾病（住院是疾病与就医行为的对撞）、按’是否点击’分层分析广告效果（点击是广告与用户兴趣的对撞）。④ M-bias 与工具变量——更复杂的情形是变量同时是混淆与对撞（M-bias），或存在未观测混淆时需用工具变量/断点回归等设计。
⚙️ 查看英文落地权衡分析 (English Systems & Trade-offs)
In ML feature engineering and causal analytics: (1) Bad Controls: Data scientists often throw all available database columns into regression. Controlling for post-treatment variables (mediators or colliders) introduces catastrophic selection bias. (2) Backdoor Criterion: A set of variables $S$ is valid for adjustment if no node in $S$ is a descendant of $X$, and $S$ blocks every backdoor path between $X$ and $Y$.
五、常见面试避坑陷阱 (Common Pitfalls & Traps)
- ⚠️ 为了’控制得更多’而误控对撞变量
- ⚠️ 控制中介后把直接效应当总效应报告
English Pitfalls:
– Conditioning on a collider (e.g. Analyzing user churn only among active users who logged in post-treatment).
– Controlling for a mediator when the business objective is to evaluate the total end-to-end effect.
六、高频深度面试追问与预测 (Follow-Up Questions)
- 为什么控制中介会低估总效应？
- What is Pearl’s Backdoor Criterion and how does it formally identify sufficient adjustment sets?
- 对撞偏差的经典例子？（天赋→录取←努力）
- What is M-bias in causal DAGs and how does conditioning on a pre-treatment variable inadvertently act as a collider?
七、知识图谱对齐 (Knowledge Graph Anchor)
- 🔗 关联底层卡片：
因果推断框架：潜在结果模型、倾向评分匹配与双重差分(Causal Inference: Potential Outcomes, PSM & DiD) - 🗺️ 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。