Tae Hyun Kim (Lowell)

d-separation

4분 읽기 #causal-inference#scm#dag

Definition

d-separation (directional separation)은 DAG에서 두 변수 집합이 세 번째 집합에 조건부로 독립인지 판단하는 graphical criterion.

Formal Definition:

DAG GG에서, ZZXXYYd-separate 한다면:

XGYZX \perp_G Y \mid Z

XXYY 사이의 모든 pathZZ에 의해 blocked.

Path Blocking Rules

Three Elementary Structures

StructureNameBlocked by ZZ?
ABCA \rightarrow B \rightarrow CChainBZB \in Z 이면 blocked
ABCA \leftarrow B \rightarrow CForkBZB \in Z 이면 blocked
ABCA \rightarrow B \leftarrow CColliderBZB \notin Z and De(B)Z=De(B) \cap Z = \emptyset 이면 blocked

Blocking Rule Summary

Path π\piZZ에 의해 blocked 되려면:

  1. Non-collider (chain or fork)의 중간 node가 ZZ에 포함
  2. Collider의 중간 node와 그 descendants가 ZZ에 미포함

Intuitive Understanding

핵심 아이디어:

d-separation = “정보 흐름의 차단”

Fork (Common Cause):
    A ← B → C

    - B가 주어지면: A, C 독립 (confounding 제거)
    - B가 주어지지 않으면: A, C 연관 (via B)

Chain (Mediation):
    A → B → C

    - B가 주어지면: A, C 독립 (mediator blocked)
    - B가 주어지지 않으면: A, C 연관 (causal path)

Collider (Common Effect):
    A → B ← C

    - B가 주어지지 않으면: A, C 독립 (default blocked)
    - B가 주어지면: A, C 연관! (collider bias)

d-separation Algorithm

Input

  • DAG GG
  • Sets XX, YY, ZZ

Procedure

For each path π between X and Y:
    blocked = False
    For each node B on path π (excluding endpoints):
        If B is a non-collider on π:
            If B ∈ Z:
                blocked = True
                break
        Else (B is a collider on π):
            If B ∉ Z and De(B) ∩ Z = ∅:
                blocked = True
                break

    If not blocked:
        return "X and Y are NOT d-separated by Z"

return "X and Y are d-separated by Z"

Faithfulness Assumption

Markov Property

PP가 DAG GG에 대해 Markov: XGYZ    XPYZX \perp_G Y \mid Z \implies X \perp_P Y \mid Z

그래프의 d-separation → 확률적 독립

Faithfulness

PP가 DAG GG에 faithful: XPYZ    XGYZX \perp_P Y \mid Z \implies X \perp_G Y \mid Z

확률적 독립 → 그래프의 d-separation

Combined (Perfect Map): XPYZ    XGYZX \perp_P Y \mid Z \iff X \perp_G Y \mid Z

Examples

Example 1: Simple Chain

X → Y → Z
  • XGZYX \perp_G Z \mid Y? Yes (Y blocks the chain)
  • XGZX \perp_G Z? No (path open)

Example 2: Fork (Confounding)

    C
   ↙ ↘
  X   Y
  • XGYCX \perp_G Y \mid C? Yes (C blocks the fork)
  • XGYX \perp_G Y? No (confounding path open)

Example 3: Collider

X → C ← Y
  • XGYX \perp_G Y? Yes (collider blocks by default)
  • XGYCX \perp_G Y \mid C? No (conditioning opens collider)

Example 4: Complex Graph

    A
   ↙ ↘
  B   C
   ↘ ↗
    D
  • AGDA \perp_G D? Check all paths:

    • ABDA \rightarrow B \rightarrow D: Open
    • Not d-separated
  • AGDBA \perp_G D \mid B?

    • ABDA \rightarrow B \rightarrow D: Blocked (B ∈ Z)
    • ACDA \rightarrow C \rightarrow D: Open
    • Not d-separated
  • AGD{B,C}A \perp_G D \mid \{B, C\}?

    • ABDA \rightarrow B \rightarrow D: Blocked
    • ACDA \rightarrow C \rightarrow D: Blocked
    • d-separated!

d-separation and Causal Inference

Back-door Criterion

Back-door Criterion은 d-separation 기반:

ZZ(X,Y)(X, Y)의 causal effect 식별에 sufficient ⟺

  • ZZXXYY 사이의 모든 back-door path를 d-separate

Adjustment Formula

d-separation을 통해 adjustment set 결정: P(Ydo(X))=zP(YX,Z=z)P(Z=z)P(Y | do(X)) = \sum_z P(Y | X, Z=z) P(Z=z)

ZZ가 back-door paths를 d-separate할 때 valid.

Extensions

m-separation (Mixed Graphs)

Bidirected edges를 포함하는 mixed graphs:

  • MAG (Maximal Ancestral Graph)
  • PAG (Partial Ancestral Graph)

σ-separation (Cyclic Graphs)

Cyclic graphs를 위한 확장:

  • Strongly Connected Components (SCC) 고려
  • d-separation으로 reduce 가능
  • DAG - d-separation이 정의되는 graph
  • Confounder - Fork structure (d-separation으로 제어)
  • Collider - Inverted fork (conditioning이 path를 open)
  • Back-door Criterion - d-separation 기반 identification
  • Markov Equivalence Class - 같은 d-separations를 가진 graphs
  • SCM - Structural Causal Model
  • Faithfulness - d-separation ↔ CI 대응 가정
  • σ-separation - Cyclic graphs 확장

References

  • Pearl, J. (1988). Probabilistic Reasoning in Intelligent Systems
  • Pearl, J. (2009). Causality
  • yaoSurveyCausalInference2021 - d-separation in causal discovery

연결 그래프