Skip to content
Scintria
ToolsGrant writerFunding
HomeLiteratureClinical trialsEvidence LabProtocol packData analysisLearnProjectsLibraryReferencesApplications
Start free
ScintriaStart free

Learn / Epidemiology / Bias, confounding and causation

Confounding, mediators, colliders and causal diagrams

14 minute read · 6-question quiz

Download this lesson as a colour Word or PDF study guide with the figures and answers. Included in paid plans; reading and quizzes stay free.Sign in

What you will learn

  • Apply the three criteria for confounding and distinguish it from a mediator and a collider

  • Draw and read a simple directed acyclic graph (DAG)

  • Choose which variables to adjust for, and which to leave alone

Confounding distorts the apparent effect of an exposure on an outcome because a third variable is related to both. It is the commonest explanation for associations that are not causal. (The arithmetic of stratified analysis is in chapter 5 of the biostatistics course; here we focus on reasoning.)

A confounder (1) is associated with the exposure, (2) is a cause or predictor of the outcome in the unexposed, and (3) is not an effect of the exposure (not on the causal pathway).

Directed acyclic graphs (DAGs)

A DAG draws variables as nodes and causal effects as arrows. It makes assumptions explicit and tells you what to adjust for. There are three basic structures.

Is there an effect?AgeConfounderCoffeeExposureHeart diseaseOutcome
Confounder: a common cause of exposure and outcome. Adjust for it.
Salt intakeExposureBlood pressureMediatorStrokeOutcome
Mediator: lies on the causal pathway. Do not adjust if you want the total effect.
TalentCause 1LooksCause 2Became a starCollider (common effect)
Collider: a common effect. Restricting or adjusting for it creates a false association between its causes.

Why colliders matter in health research

Suppose two diseases are unrelated in the population, but either one can lead to hospital admission. Among hospital patients, having one disease makes the other less likely to be the reason for admission, so the two appear negatively associated. This is Berkson's bias, a collider effect. Likewise, studying only 'survivors' or only 'people who were tested' can open collider paths. The lesson: conditioning on a consequence of the exposure or the outcome can create bias.

A decision rule for adjustment

For each candidate variable ask:

  1. Does it cause (or share a common cause with) both the exposure and outcome? Adjust (confounder).
  2. Is it caused by the exposure and does it lead to the outcome? Do not adjust if you want the total effect (mediator); adjust only to study direct effects.
  3. Is it caused by both exposure and outcome (or their causes)? Do not adjust (collider).
  4. Does it only cause the outcome (not exposure)? Adjusting can improve precision, harmless.
  5. Does it only cause the exposure? Adjusting can hurt precision; usually leave out.

Does vitamin A supplementation reduce child deaths?

In an observational study, children who received vitamin A had lower mortality. But health-seeking parents (who are wealthier, more educated, near clinics) both get vitamin A and have healthier children. That is confounding. The authors adjust for wealth, education, distance and prior illnesses (confounders). They do not adjust for 'diarrhoea in the past month', which might be an effect of vitamin A status and a cause of death (mediator). Randomised trials provided the definitive answer, because randomisation breaks the link between the parents' characteristics and who gets the supplement.

Methods to control confounding

Design: randomisation, restriction, matching. Analysis: stratification, standardisation, regression adjustment, propensity scores, instrumental variables. None controls for factors that were not measured: this is residual confounding. Always ask what important confounder was missed.

'Adjusted for' is not a magic phrase

Adjusting for a poorly measured confounder (e.g. income by crude categories) leaves much confounding. Adjusting for everything available can introduce collider and mediator problems. Think first, then adjust.

Summary

  • A confounder is related to exposure and outcome and is not on the causal pathway.

  • A mediator lies between exposure and outcome; a collider is caused by both.

  • DAGs make assumptions explicit and guide adjustment.

  • Adjust for confounders; do not adjust for mediators (for total effects) or colliders.

  • Randomisation controls unmeasured confounders; observational studies leave residual confounding.

Check your understanding

Question 1
1

In a DAG, a variable with arrows pointing into it from both exposure and outcome is a:

Question 2
2

Adjusting for a collider can:

Question 3
3

Blood pressure on the path from salt to stroke is a:

Question 4
4

Berkson's bias arises because:

Question 5
5

Which design best controls unmeasured confounders?

Question 6
6

'Residual confounding' means:

Draft study guide. Spotted an error or something unclear? Tell us at hello@scintria.com. This is educational material, not clinical or regulatory advice.

Selection bias and information biasFrom association to causation