Confounding

Confounding occurs when linked variation makes the separate effects of two or more factors impossible to distinguish from observed data alone.

Confounding variable: A variable associated with both an exposure and an outcome that can distort their observed relationship. It creates confounding when its variation is entangled with the factor being studied.
Causal inference: The study of methods for estimating cause-and-effect relationships from data. Confounding is a central obstacle to identifying causal effects.
Selection bias: Bias caused when inclusion in a study or analysis depends on characteristics related to exposure or outcome. It arises through selective inclusion, rather than necessarily through linked exposure and confounder variation.
Smoking and lung cancer: A well-established causal relationship between tobacco smoking and increased risk of lung cancer. Early observational comparisons faced potential confounding by factors such as age and occupation.
Unmeasured confounding: Confounding by factors that were not measured or adequately represented in an analysis. It can remain after adjustment for every recorded covariate.
Omitted-variable bias: Bias in an estimated relationship caused by leaving out a relevant variable associated with included predictors. It is a regression form of confounding when the omitted factor tracks a studied predictor.
Causal diagram: A graph representing assumed causal relationships among variables. The diagram reveals paths that link an exposure to an outcome through other factors.
Collider bias: A spurious association induced by conditioning on a variable caused by two other variables. Adjusting for a collider can create bias even when adjustment is intended to address confounding.
Healthy user bias: Bias arising when people who choose a treatment also tend to engage in health-promoting behaviors. Those behaviors can confound comparisons of treatment users with nonusers.
Residual confounding: Confounding that persists after adjustment because a factor is measured or modeled imperfectly. It explains why adjustment does not guarantee removal of a confounder's influence.

Linked from 46 pages