d-Separation and Conditional Independence
You'll learn: How to use formal d-separation rules to read conditional independence relationships directly from a DAG.
What Is d-Separation?
d-separation (directional separation) is a graphical criterion that tells you whether two variables are conditionally independent given a set of observed variables—just by looking at the DAG structure. If two nodes are d-separated given some conditioning set, they are conditionally independent; no statistical association flows between them once you account for the conditioning variables.
Think of it like checking whether water can flow between two points in a pipe network. If you've closed certain valves (conditioned on certain variables), can the flow still reach from A to B?
The Three Path Types
A path between two variables can be blocked or open depending on what you condition on:
- Chain (A → M → B): Open by default. Blocked if you condition on M (the mediator).
- Fork (A ← C → B): Open by default. Blocked if you condition on C (the confounder).
- Collider (A → C ← B): Blocked by default. Opens if you condition on C or its descendants.
How to Use d-Separation
To determine if X and Y are conditionally independent given set Z:
- Trace all paths between X and Y
- Check if every path is blocked by Z
- If yes: X and Y are d-separated (conditionally independent)
- If no: they remain dependent
This rule helps you decide which variables to control for—and critically, which not to control for (like colliders).
Key Takeaway: d-separation provides formal rules to read conditional independence from DAG structure: chains and forks block when you condition on the middle node; colliders block by default but open when conditioned upon.