Confounding

Confounding is a mixing of effects: an apparent association between an exposure and an outcome is distorted because a third factor — the confounder — is a common cause of both. Unless it is accounted for, confounding can make a harmless exposure look harmful, hide a real effect, or exaggerate one. Controlling confounding is one of the central tasks of observational epidemiology.

Definition

Confounding is the distortion of an exposure-outcome association that arises when a third variable is a common cause of both the exposure and the outcome (and is not on the causal pathway between them), so that the crude association mixes the effect of the exposure with the effect of that variable.

Scope

The entry covers what makes a variable a confounder, how confounding is recognised using causal reasoning and directed acyclic graphs, and the main strategies for controlling it by design and analysis. It also distinguishes confounding from effect modification and from mediation. It is a methodological reference, not clinical guidance.

Core questions

• Is a candidate variable a common cause of both exposure and outcome?
• Is the variable on the causal pathway (a mediator) rather than a confounder?
• What set of variables must be adjusted for to remove confounding?
• Could residual or unmeasured confounding still explain the association?

Key concepts

• Common cause
• Confounder
• Directed acyclic graph (DAG)
• Exchangeability
• Backdoor path
• Adjustment, stratification, and matching
• Residual and unmeasured confounding
• Confounding versus mediation

Mechanisms

Under a potential-outcomes view, confounding is a failure of exchangeability: the exposed and unexposed groups differ in their background risk of the outcome for reasons other than the exposure. In causal-diagram terms, a confounder opens a non-causal 'backdoor' path connecting exposure and outcome through a common cause; blocking that path — by conditioning on an appropriate set of variables — removes the confounding. A variable on the causal pathway (a mediator) is not a confounder, and adjusting for it can introduce its own bias. Control can be exercised by design (randomisation, restriction, matching) or by analysis (stratification, standardisation, regression adjustment, and methods such as propensity scores). Because adjustment can only handle measured confounders, unmeasured and residual confounding remain a limitation of observational estimates.

Clinical relevance

Confounding is a primary reason that observational associations are not automatically causal, and judging how well a study handled it is central to appraising evidence about exposures and disease. The concept explains how evidence is interpreted; it is not itself advice about diagnosis or treatment for any individual.

Epidemiology

Concern about confounding pervades observational research across cohort, case-control, and cross-sectional designs. It motivates randomised trials (which, on average, balance both measured and unmeasured confounders) and the growing use of explicit causal diagrams and quantitative bias analysis to reason about what must be adjusted for.

History

Awareness that a third factor can mix effects is old, but precise definitions emerged in the late twentieth century. Greenland and Robins (1986) framed confounding through exchangeability and the counterfactual comparison the study is trying to approximate, and the causal-diagram framework of Greenland, Pearl, and Robins (1999) gave a graphical criterion for identifying confounders and choosing adjustment sets. These developments separated confounding cleanly from effect modification and from selection bias.

Debates

How should confounders be selected for adjustment?

Older practice leaned on statistical criteria (for example, change-in-estimate or significance of associations), whereas causal-diagram approaches argue that adjustment sets should be chosen from prior subject-matter knowledge of the causal structure to avoid adjusting for mediators or colliders.

Key figures

• Sander Greenland
• James Robins
• Judea Pearl
• Olli Miettinen

Seminal works

• greenland-robins-1986
• greenland-pearl-robins-1999
• maldonado-greenland-2002

Frequently asked questions

What is the difference between a confounder and a mediator?

A confounder is a common cause of both exposure and outcome and lies off the causal pathway; a mediator lies on the pathway from exposure to outcome. Adjusting for a confounder reduces bias, but adjusting for a mediator can introduce it.

Does randomisation remove confounding?

Randomisation tends, on average and especially in large trials, to balance both measured and unmeasured common causes between groups, which is why it controls confounding in a way that adjustment of observational data cannot fully match.

Is confounding the same as effect modification?

No. Confounding is a distortion to be removed; effect modification is genuine variation in the effect across subgroups and is a property of the relationship, not an error.