ConceptsGlossary

Glossary

Short definitions of the terms these pages use, grouped by topic. Each group links to the page that puts the terms to work.

Causal ideas#

Treatment and outcome: the variable you change and the variable you measure.
Confounder: a common cause of the treatment and the outcome. Left unadjusted, it makes them move together without one causing the other.
Mediator: a variable on a causal path from treatment to outcome.
Collider: a node two arrows point into (X -> Z <- Y). It blocks a path until you condition on it.
Intervention, do(X): setting X yourself rather than observing it. On a graph, it removes every arrowhead into X.
Overlap (positivity): every kind of unit could receive either treatment, so each has someone to be compared with.
Selection bias: distortion from keeping only the rows where some condition held. Conditioning on it can make its causes look related.

Mark: the symbol at one end of an edge: tail, arrowhead or circle.
Glyph: an edge written as a string, such as A -> B, A <-> B or A o-> B.
DAG: directed acyclic graph. Every edge has a direction, and there are no cycles.
ADMG: a DAG that adds <-> edges for hidden common causes.
CPDAG: the graphs that data cannot tell apart, drawn as one graph. Directed edges hold in every member; -- edges point either way.
MAG: a graph over the measured variables that keeps what hidden causes and selection imply about independence.
PAG: the MAGs that data cannot tell apart, drawn as one graph. A circle marks an end that differs between them.
Kind: which of these readings a graph uses; mixed means none is checked.
Latent node: a cause in the graph that is not a column in the data.
Markov equivalence class: the graphs that imply the same independences.
v-structure: a collider whose two parents are not adjacent. Data can orient it.
d-separation, m-separation: the graph test for whether two variables are independent given a set of others.
Graph format: the version tag to_dict writes, ergodic.graph/1 today.

Conditional-independence test: a test of whether two variables are independent given others. PC and FCI are built on one.
Score: a number that rates how well a graph fits the data, such as BIC. GES and exact search maximize one.
Faithfulness: every independence in the data comes from the graph, not from effects cancelling out.
Domain knowledge: constraints you know before looking at data: required or forbidden edges, tiers, roots and sinks.

Identifiable: the effect can be computed from the distribution of the measured variables, given the graph.
Estimand: the formula identification returns, written in measurable quantities.
Adjustment set: variables to condition on so that every non-causal path between treatment and outcome is blocked.
Front-door: identifying a confounded effect through mediators that the hidden confounder does not reach.
Instrument: a variable that moves the treatment and reaches the outcome only through it.
Null strategy: no directed path runs from treatment to outcome, so the effect is zero by the graph.
ID algorithm: the general procedure that decides identifiability for any single DAG or ADMG; a hedge is its proof that an effect is not identifiable.
Natural direct and indirect effects: the part of an effect that does not go through a mediator, and the part that does.

Average treatment effect (ATE): the mean outcome if everyone were treated, minus the mean if no one were.
ATT: the same contrast, averaged over the treated units only.
CATE: the average effect for units that share given features.
Learner: a model in a nuisance slot, with fit and predict.
Nuisance slot: a model an estimator needs but does not report, such as the outcome or the propensity model.
Propensity: the probability of being treated, given the confounders.
AIPW: augmented inverse-probability weighting. It combines an outcome model and a propensity model and stays consistent if either one is right.
DML: double machine learning. It removes what the confounders predict from the treatment and the outcome, then regresses one residual on the other.
Cross-fitting: fitting nuisance models on some folds and predicting on the others, so no row is predicted by a model trained on it.
Refutation: rerunning an estimate under a perturbation it should survive or fail.

Difference-in-differences: compares the change in treated units with the change in comparison units; rests on parallel trends.
Synthetic control: a weighted mix of untreated units that tracks the treated unit before treatment.
Interrupted time series: compares a series after a break with the trend it followed before.

Welt: Ergodic's hosted service for tabular foundation models, causal discovery included.
Native graph type: the kind of graph a Welt model declares it returns. ergodic converts a declared DAG directly. It refuses cdfm and avici, whose graphs are not declared DAGs, unless you pass an explicit decoder= that decodes their edge scores locally. See Decode a DAG from edge probabilities.
Fit budget: max_fits, the most paid fits a WeltLearner and its copies may make.