Confounding

Confounding happens when the association seen between an exposure and an outcome is partly or wholly produced by another factor, called a confounder. To act as a confounder, a factor must influence the outcome on its own, be unevenly spread between the groups being compared, and not be a step on the causal path from exposure to outcome. In a classic example, coffee drinkers seemed to have more lung cancer, mainly because they were more often smokers. Smoking was the confounder.

Confounding can make a harmless exposure look harmful, make a real effect disappear or change its size. It is a constant threat in observational research such as cohort and case–control studies, where people are not assigned to their exposures. Researchers limit it at the design stage by randomizing, restricting who can take part or matching groups on key traits, and at the analysis stage by stratifying the data or adjusting for measured confounders in statistical models.

Randomization is the strongest defense because, in a large enough trial, it balances unknown confounders as well as known ones. A common pitfall in drug studies is confounding by indication, in which sicker patients are more likely to receive a treatment and so do worse for reasons unrelated to the drug. Some residual confounding from unmeasured factors may remain even after careful adjustment.