Causal AI glossary

Plain definitions of the terms of causal discovery and causal inference: DAG, confounder, collider, counterfactual, intervention, CPDAG, PAG, faithfulness and more.

Graphs

Causal graph: a diagram of variables (nodes) and direct causal links (arrows) between them.
DAG (directed acyclic graph): a causal graph in which every link has a direction and no path leads back to where it started.
CPDAG: a partially directed graph standing for all the DAGs the data cannot tell apart; links whose direction differs among them are left undirected. PC, GES, BOSS and GRaSP return one.
PAG (partial ancestral graph): FCI's output, whose link ends carry marks (arrow, tail or circle) that allow for hidden causes.
Skeleton: the graph's links without their directions.
Path: a chain of links from one variable to another. A causal path follows the arrows.
Parent / child: A is a parent of B when A → B; B is then a child of A.

Patterns

Hand-drawn labelled causal graph: a confounder causes both X and outcome Y, X affects Y through a mediator, and X and Y both cause a collider — Confounder, mediator and collider in one graph

Confounder (common cause): a variable that causes two others and so makes them move together.
Collider (common effect): a variable caused by two others. Accounting for it creates a false link between them.
Mediator: a variable on the path from cause to effect (A → M → B).
Hidden (latent) variable: a cause that is not in the table.
Selection bias: a distortion that comes from keeping only rows that pass a filter related to the variables studied.
Reverse causation: the effect is mistaken for the cause.

Methods and assumptions

Causal discovery: learning a causal graph from data.
Constraint-based method: builds the graph from independence tests (PC, FCI).
Score-based method: searches for the graph that best explains the data under a score (GES, BOSS, GRaSP).
Conditional independence: A and B are unrelated once C is known.
Faithfulness: the assumption that every independence in the data comes from the graph's structure, not from effects that cancel out.
Causal sufficiency: the assumption that no hidden variable causes two measured ones.
Non-Gaussian noise: variation that is not bell-shaped; it lets LiNGAM tell directions apart.
Additive noise model: an effect equals a function of its causes plus independent noise.
BIC (Bayesian information criterion): a score that rewards fit and penalises extra links.
Fisher-z test: an independence test for roughly linear, bell-shaped data.

Effects and decisions

Intervention: setting a variable to a value from outside, as in an experiment, rather than observing it.
Causal effect: how much an outcome changes when a cause is changed by intervention.
Direct / indirect effect: the part of an effect along the direct link versus through mediators.
Counterfactual: what would have happened to a case had something been different.
What-if scenario: an estimate of an outcome under a planned change.
Randomised experiment (A/B test): assigning a change by chance so that nothing else can explain the difference.
Observational data: data recorded as things happened, without assigning any change.

Reading results

Undirected link: a link the method found but could not orient from the data.
Edge weight: the size of a link's effect in a linear model, in the units of the columns.
Confidence (stability): how consistently a link appears when the data is resampled; only some methods measure it.
Consensus: the links that enough methods agree on.