Confounders, Colliders, and Mediators

Three Core Roles: Definitions

In causal diagrams (DAGs), each variable occupies a distinct structural position in the relationship between exposure X and outcome Y. A confounder C has arrows pointing to both X and Y; it is a common cause. A mediator M sits on the path from X to Y: X affects Y through M. A collider K receives arrows from both X and Y; it is a common effect of the two variables. These three structures generate fundamentally different inferential problems and require different analytic strategies.

Adjustment Logic: When to Control and When Not To

For a confounder, including it in a regression or matching on it removes confounding bias. For a mediator, adjusting blocks the portion of the total causal effect that flows through it; this is an error if you want the total effect. For a collider, adjustment opens a previously closed path and creates an association that does not exist in the population; this is the mechanism behind Berkson bias. The guiding rule: close confounders, leave mediators open when estimating total effects, and never adjust for colliders.

Concrete Example: Education, Income, and Health

Consider the relationship between education X and health Y. Socioeconomic background affects both education and health, making it a confounder that should be controlled. Income M is a pathway through which education influences health; including it in the model blocks part of the effect if you want the total causal effect. Now suppose you study only hospitalized individuals: if hospitalization is a common effect of both education and health status, it is a collider. Restricting the sample to this condition opens a spurious association between education and health that does not exist in the general population.

Common Pitfalls and Good Practice

The most common error is deciding a variable's role by looking at statistical associations. Both a collider and a confounder may correlate with X and Y; the distinction can only be made from the causal structure in a DAG. A second error is automatically adjusting for mediators, which underestimates the causal effect and produces biased estimates. Good practice: clarify the research question, draw a DAG based on prior subject-matter knowledge, read each variable's role from that diagram, and only then make analytic decisions. Adjustment choices made without a DAG are little more than guesswork.

Key terms

Confounder

A common cause of both exposure and outcome; its presence creates confounding bias.

Mediator

A variable on the causal path from exposure to outcome; part of the mechanism.

Collider

A common effect of two variables; adjusting for it induces spurious association.

Directed Acyclic Graph (DAG)

A graph encoding assumed causal relationships among variables without cycles.

Berkson Bias

Selection bias arising in clinical samples from conditioning on a collider variable.

Further reading