Causal Structure Discovery from Correlational Data (IC*, PC/TETRAD)

Description

Uses patterns of conditional independence among observed variables to build a partially directed causal graph showing definite, possibly confounded and ambiguous links, as a hypothesis generator.

Constraint-based causal discovery reads causal structure from which variables become independent when others are held constant. The core inference: if A and C are each associated with B but independent of each other, and become dependent once B is conditioned on, then A and C both point into B (an unshielded collider). Starting from all associated pairs, the algorithm removes links between variables that some conditioning set renders independent, orients colliders, and propagates further arrowheads by rule. Pearl's IC* allows for latent common causes, so its output distinguishes genuine causal links, links that may reflect hidden confounders, and fully ambiguous links. The result is not one model but a class of equivalent models; known temporal order fixes many directions and speeds the search. Loehlin and Beaujean compare it to exploratory factor analysis: valuable for generating and constraining hypotheses, not a replacement for confirmatory modelling or experiments.

Attributes

Loss type
Unmigrated
Effort
Weeks

Steps

  1. Select variables measured on the same units, ideally with known temporal order (for example signup attributes before early usage before retention).
  2. Test conditional independence for each pair given subsets of the other variables, and connect only pairs that no conditioning set makes independent.
  3. Find unshielded triples (A and C linked to B but not to each other) where A and C are dependent given B, and orient both arrows into B.
  4. Apply propagation rules to add further arrowheads, using temporal order as background knowledge.
  5. Read the output as definite causes, possible latent confounding and ambiguous links, and list the equivalent models it implies.
  6. Treat the graph as hypotheses: check stability across subsamples and test key links with experiments or quasi-experiments.

Sources

  • Latent Variable Models (Loehlin & Beaujean, 2017)
  • Bayesian Networks: An Introduction (Koski & Noble, 2009)

Also known as

  • Automatic construction of path diagrams
  • Inductive causation algorithm
  • Constraint-based causal search
  • TETRAD causal search
  • PC algorithm
  • Causal discovery

Categories

  • Analytics & Behaviour
  • Experimentation & Causal Inference