M1-057M1: Mathematics & Statistics FundamentalsCausal InferenceEasy
⚡Flashcard Drill➔Mastery:

Explain Confounders, Mediators, and Colliders in Causal DAGs and Which Must or Must Not Be Controlled

🎯 Active Recall Self-Check (Draft First, Verify After)
Login-free practice tracker: recall core answer, audit pitfalls, and log review status.
Step 1 · Draft your concise answer & assumptions
Step 2 · Check Pitfalls & Follow-up Inquiries
Next: M1-058 →
📐 Mathematical Definition
confounder:X→T, X→Y;collider:T→X←Y\text{confounder}: X\to T,\ X\to Y;\quad \text{collider}: T\to X\leftarrow Y
⚡ One-Sentence Core Answer
Confounders (forks X←Z→YX \leftarrow Z \to Y) cause spurious correlation and MUST be controlled; Mediators (X→M→YX \to M \to Y) transmit mechanisms and should NOT be controlled for total effect; Colliders (X→C←YX \to C \leftarrow Y) induce artificial association when controlled and MUST NOT be controlled.

📌 Key Takeaways

  • •
    Confounder (ZZ): Common cause of both treatment XX and outcome YY; creates non-causal backdoor paths that MUST be blocked/conditioned on.
  • •
    Mediator (MM): Intermediate step along the causal pathway X→M→YX \to M \to Y; controlling for MM blocks the indirect causal effect, leaving only direct effect.
  • •
    Collider (CC): Common effect of both treatment XX and outcome YY; conditioning on a collider OPENS a spurious association path (Berkson's Paradox).

📐 Mathematical Derivations

DAG structural mechanics via d-separation: (1) Fork X←Z→YX \leftarrow Z \to Y: XX and YY are dependent: P(X,Y)≠P(X)P(Y)P(X, Y) \ne P(X)P(Y). Conditioning on ZZ blocks the path: (X⊥Y∣Z)(X \perp Y \mid Z). Hence, adjusting for confounders isolates the true causal arc X→YX \to Y. (2) Chain X→M→YX \to M \to Y: Conditioning on MM blocks the transmission: (X⊥Y∣M)(X \perp Y \mid M). Doing so eliminates the primary causal mechanism. (3) Collider X→C←YX \to C \leftarrow Y: Unconditioned, XX and YY are marginally independent: (X⊥Y)(X \perp Y). However, conditioning on CC induces correlation: (X⊥̸Y∣C)(X \not\perp Y \mid C). For example, let XX = Talent, YY = Hard Work, CC = College Admission (C=X+YC = X + Y). Among admitted students (C=1C=1), knowing someone is not talented implies they must be exceptionally hardworking, inducing artificial negative correlation.

🏭 Production 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 SS is valid for adjustment if no node in SS is a descendant of XX, and SS blocks every backdoor path between XX and YY.
⚠️ Common Interview 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.
🎯 Interviewer Follow-up Inquiries
  • ?
    What is Pearl's Backdoor Criterion and how does it formally identify sufficient adjustment sets?
  • ?
    What is M-bias in causal DAGs and how does conditioning on a pre-treatment variable inadvertently act as a collider?
⚡
Test muscle memory on this topic via 3-Layer Flashcards
Active Recall: Mentally formulate Prompt ➔ Math Derivation ➔ Benchmark Implementation in 3 minutes.
🎯 Drill This Card Now➔
📚

Associated Knowledge Base Guides & Mindmaps

← PreviousM1-056: Define Causal Effects and the Potential Outcomes Framework (Rubin Causal Model)📋 Back to BankNext →M1-058: Explain Propensity Score Matching (PSM) and Inverse Probability Weighting (IPW) for Observational Causal Inference