D-separation

Description

A graphical rule that reads off a DAG which variables should be statistically independent given others, enabling tests of causal assumptions against data.

D-separation determines whether a set of variables blocks every path between two other variables in a DAG. A path is blocked if it contains a non-collider that is conditioned on, or a collider that is not conditioned on (nor any of its descendants). If all paths are blocked, the two variables are d-separated and must be conditionally independent in any data generated by that graph. This gives two uses for UX research: checking a proposed causal model by testing its implied independencies against survey or log data, and understanding why controlling for a variable opens or closes bias. Rigour comes from computing implications mechanically with software and testing them with appropriate conditional independence tests.

Attributes

Loss type
Unmigrated
Stage
Definition
Effort
Hours

Steps

  1. Encode the DAG in a tool such as DAGitty or the Python package pgmpy.
  2. List all implied conditional independencies the graph makes, which the tool generates automatically.
  3. Test each implication in data with partial correlations or conditional independence tests, correcting for multiple tests.
  4. Treat strong violations (e.g. partial correlation above 0.1 with a tight interval) as evidence the graph is missing an edge or variable.
  5. Revise the DAG and repeat, documenting every change and its evidential basis.

Sources

  • DAGs+SEM Playbook (Vivek Gopinath)
  • Causal Inference: What If (Miguel Hernán & James Robins)
  • Statistical Rethinking (Richard McElreath)
  • Probabilistic Reasoning in Intelligent Systems (Judea Pearl)
  • DAGs+SEM Playbook (Gopinath)
  • Causal Inference (Hernán & Robins)
  • Statistical Rethinking (McElreath)
  • The Inquiry Series, Ch 46 — Directed Acyclic Graphs — Making Assumptions Explicit

Also known as

  • Directional separation
  • Graphical independence criterion

Categories

  • Information Theory, Physics & Mathematics