Causal Inference

Overview

Causal inference asks what would happen to an outcome if a treatment, policy or exposure were changed while other relevant conditions were held comparable. Because the same unit cannot be observed simultaneously under two alternatives, researchers combine explicit assumptions, study design and statistical models to estimate counterfactual contrasts from incomplete evidence.

Technical foundations

The potential-outcomes framework defines a unit's response under each treatment level and expresses an average treatment effect as a contrast between those counterfactual quantities. Graphical causal models instead represent variables and directed structural relations, using d-separation to determine which conditional independences follow from a graph. The two formalisms meet through assumptions such as consistency, exchangeability and positivity. Identification is distinct from estimation: first one proves that the target effect can be written as a functional of the observed-data distribution, and only then chooses an estimator for that functional.

How it works

A causal question defines the intervention, target population, outcome and time horizon. Random assignment can balance known and unknown causes on average. In observational data, adjustment, matching, weighting, instrumental variables or quasi-experimental designs attempt to reconstruct a valid comparison by blocking non-causal paths without conditioning on variables that introduce selection bias.

Measurement and research methods

Randomised trials estimate intention-to-treat effects under adherence and missingness rules fixed in advance. Observational studies use standardisation, inverse-probability weighting, matching or doubly robust estimators after defining an admissible adjustment set. Instrumental-variable, regression-discontinuity and difference-in-differences designs exploit specific sources of quasi-random variation, each with separate exclusion, continuity or parallel-trend assumptions. Diagnostics examine overlap, covariate balance and model dependence, while negative controls and placebo tests probe selected failure modes. Confidence intervals address sampling variation but do not automatically include uncertainty about the causal graph or unmeasured confounding.

Key ideas

• Association becomes causal evidence only through assumptions tied to a defensible design.
• A directed acyclic graph can expose confounders, mediators and colliders before modelling begins.
• Sensitivity analysis should quantify how unmeasured bias could change the conclusion.

Current research frontier

Current research combines flexible machine learning with orthogonal scores, targeted learning and cross-fitting so nuisance models can be complex without dominating treatment-effect inference. Heterogeneous-effect methods estimate how benefits and harms vary across groups, but subgroup discovery requires multiplicity control and external validation. Transportability methods formalise whether trial results apply to a new population, while causal discovery searches for graph structure under assumptions rarely satisfied perfectly. Important frontiers include interference in networks, dynamic treatment regimes, longitudinal time-varying confounding and sensitivity analyses that report how strong an omitted common cause must be to reverse a decision.

Why it matters

Reliable causal estimates guide medicine, public policy, education and product decisions where prediction alone cannot identify the effect of an action. The framework also makes disagreements auditable by separating data-supported quantities from assumptions.

Limits and open questions

Causal effects may not transport across populations, periods or implementation conditions. Positivity violations, interference between units, measurement error and unmeasured confounding can make an estimand weakly identified or impossible to recover from the available data.