Confounding
- Bias
A third factor, linked to both exposure and outcome, creates or hides an association that is not causal.
What it is
Confounding occurs when the association between an exposure and an outcome is mixed with the effect of a third factor. To be a confounder, that factor must be associated with the exposure, be a cause (or marker of a cause) of the outcome in its own right and not lie on the causal path between them. Confounding shows up when the crude estimate differs from the adjusted or stratified one, and it is controlled at design or in the analysis.
How it arises
- The factor (for example, age) is more common among the exposed.
- The same factor raises the risk of the outcome on its own.
- When exposed and unexposed are compared, part of the difference is due to the factor, not the exposure.
Worked example
- Cohort of 2 000 adults followed 5 years: 800 with grey hair (600 aged 60 or more) and 1 200 without (200 aged 60 or more).
- Heart attack risk: 10% at 60 or more and 2% under 60, with or without grey hair.
- Grey hair: 600 × 0.10 + 200 × 0.02 = 60 + 4 = 64 heart attacks; 64/800 = 8%.
- No grey hair: 200 × 0.10 + 1 000 × 0.02 = 20 + 20 = 40 heart attacks; 40/1 200 ≈ 3.3%.
- Crude RR = 8% / 3.3% ≈ 2.4.
- By age stratum: 60 or more, 10% versus 10% (RR = 1.0); under 60, 2% versus 2% (RR = 1.0).
- Reading: grey hair does not cause heart attacks; age, linked to both, confounded the crude association.
Confounding versus other third-factor situations
| Situation | What it is | What to do |
|---|---|---|
| Confounding | The factor distorts the overall association | Adjust or stratify and report the adjusted measure |
| Effect modification | The real effect changes with the factor | Report each stratum separately |
| Intermediate variable | The factor lies on the causal path | Do not adjust for it when the total effect is wanted |
How to prevent or control it
- Design: randomisation, restriction or matching.
- Analysis: stratification, the Mantel-Haenszel method or multivariable models.
- Even after adjustment, residual confounding remains from poorly measured or unknown factors.
Common pitfalls
- Adjusting for an intermediate on the causal path: part of the real effect is erased.
- Thinking a well-adjusted observational study removes all confounding.
- Mixing up confounding and effect modification: with confounding, the strata agree with each other and differ from the crude estimate.
Test yourself
Think of the answer, then tap the question to see it.
What three conditions must a confounder meet?
Be associated with the exposure, be associated with the outcome in its own right, and not be an intermediate step on the causal path.
How is confounding detected in the analysis?
By comparing the crude estimate with the adjusted or stratified one: if they differ meaningfully, there is confounding.
Which design best controls confounding by unknown factors?
The randomised controlled trial, because chance also spreads unmeasured factors evenly.
Connections
Educational content on epidemiological and statistical methods; the examples use made-up numbers to teach the calculation. Not a clinical or legal guide: current Colombian rules are in the official sources.
Sources: Confounding | Catalog of Bias (Catalogue of Bias Collaboration, 2018); Chapter 25: Assessing risk of bias in a non-randomized study (Cochrane, 2024); Principles of Epidemiology | Lesson 1 - Section 7 (Centers for Disease Control and Prevention, 2012).
Spotted a mistake?We are always improving for the community. Tell us what is wrong and our team will review it.Report a mistake
Representative model, for educational purposes only.