Confounding Variable

A confounding variable $Z$ distorts the observed association between an exposure or treatment $X$ and an outcome $Y$ when $Z$ is causally or statistically associated with both $X$ and $Y$, and is not itself a consequence of $X$ (a variable on the causal pathway between $X$ and $Y$, a mediator, should not be treated as a confounder to adjust away). When a true confounder is left unaccounted for, an observed association between $X$ and $Y$ may be entirely or partially spurious, reflecting the shared influence of $Z$ rather than a genuine causal effect of $X$ on $Y$; standard remedies include randomization (which balances confounders, measured and unmeasured, across groups in expectation), matching or stratification on known confounders at the design stage, and statistical adjustment (including the confounder as a covariate in a regression model) at the analysis stage, though the latter can only address confounders that were actually measured, leaving the possibility of unmeasured confounding an ever-present limitation of purely observational study designs. This general survey-and-study-design concept, a real-world third variable distorting an observed association between an exposure and an outcome, is distinct from confounding in the technical experimental-design sense, where specific higher-order interaction effects in a factorial design are deliberately made statistically indistinguishable from each other or from block effects as an efficiency trade-off, an intentional structural feature of certain designs rather than an unwanted, real-world source of bias.

• Randomization is the strongest available protection against confounding, because it balances not just known, measured confounders but also unmeasured and even unknown ones across treatment groups in expectation, this is precisely why randomized controlled trials are considered the strongest design for causal inference and why purely observational studies remain perpetually vulnerable to unmeasured confounding.
• A confounder must be distinguished from a mediator, a variable that lies on the causal pathway between the exposure and the outcome (X causes the mediator, which in turn causes Y), statistically adjusting for a mediator, rather than a true confounder, can mask or bias away part of the very causal effect a study is trying to measure, a common and consequential analytical mistake.
• This general concept of a confounding variable, a real third factor distorting an exposure-outcome relationship in an observational or experimental study, is distinct from confounding (experimental design), the technical device in factorial experiments where specific higher-order interaction effects are deliberately made indistinguishable from each other as a calculated efficiency trade-off, the two share a name and a mathematical flavor (inseparability of effects) but arise from entirely different sources, one unwanted and real-world, the other intentional and structural.