MerQuR+Statistics Glossary

Statistics Glossary → Confounding variable

Confounding variableconcept

A third variable that is related both to the exposure or group under study and to the outcome, and so distorts the relationship between them.

A confounder can mimic, mask or even reverse the effect of an exposure on an outcome: because both depend on a common cause, they can be associated without any real effect. To count as a confounder, a variable must be related to the outcome, related to the exposure and not be a consequence of the exposure (a mediator on the causal path). Simpson's paradox shows that ignoring a confounder can even flip the direction of an association.

Confounding is controlled at the design stage by randomization, matching or restriction, and at the analysis stage by stratification (for example Mantel-Haenszel), regression adjustment or propensity score methods. These methods can only handle confounders that were measured. Adding mediators or colliders (variables affected by both exposure and outcome) to a model as if they were confounders can create new bias.

An everyday example

Lung cancer turns out to be more common among coffee drinkers. But coffee drinkers are also more likely to smoke, and smoking causes lung cancer. When smokers and non-smokers are examined separately, the link between coffee and cancer largely disappears in both groups: smoking is the confounder.

In MerQur

In the Cochran-Mantel-Haenszel test the confounder is entered as the stratum; the Mantel-Haenszel common odds ratio pooled across strata gives the association adjusted for the confounder, and the Breslow-Day test checks whether the association is homogeneous across strata.

Analyses where this concept appears

Cochran-Mantel-Haenszel (CMH) Test

Related concepts

Stratified analysis (stratum variable)Covariate