Confounding

A third variable that influences both the exposure (or treatment) and the outcome being studied, producing an association between them that is not causal. If smoking makes someone more likely to take a drug and independently raises the risk of the injury the drug is being studied against, a study that never accounts for smoking will find the drug and the injury linked even if the drug does nothing at all. The word "control," used in the sense of "controlling for a variable," refers to removing this kind of distortion — the worked example and the term itself are from ramos-controlling-variables.

Two ways to control for it

• Experimental control (randomization). Assign the treatment at random. In expectation, a confounder is distributed evenly across the treatment and control arms, so it stops predicting which arm a participant is in and can no longer manufacture an association. This is control exercised before the data exists.
• Statistical control (adjustment). When the researcher did not assign the treatment — an observational study built from existing records — a statistical model instead adjusts for the suspected confounder after the fact (e.g. including age or sex as a covariate). This is control exercised on data whose generation the researcher had no say in, and it only works for confounders the analyst thought to measure and include.

Stratification is randomization's version of blocking

Randomizing separately within subgroups that share a trait (age band, baseline blood sugar) prevents one arm from being loaded with people who differ on that trait by chance, and tightens the resulting effect estimate. This is the same design larsen-marx-mathematical-statistics teaches under randomized block designs (ch. 13: block on a nuisance variable, then randomize within each block) — ramos-controlling-variables gives it a clinical trial to motivate it rather than an agricultural or industrial one.

Two ways adjustment can fail, in opposite directions

• Too few controls. Omitting a real confounder leaves a spurious association standing — the textbook case is a self-selection effect (healthier people choosing to take a supplement) being mistaken for the supplement's own effect. Findings built this way often do not survive a randomized trial.
• Too many controls. Adjusting for variables that are not confounders makes the estimate less precise without removing any bias. Worse, adjusting for a collider — a variable that is itself an effect of both the exposure and the outcome, rather than a cause of either — can introduce a spurious association that was not there to begin with. Controlling for a variable is not uniformly safe; it depends on that variable's causal role, which the statistics alone does not reveal.

Where it sits in the corpus

mathematical-statistics teaches estimation and testing once the data-generating process is already specified as an assumption (a random sample, an i.i.d. draw). Confounding is upstream of that: it is about whether the data-generating process itself entangles the thing being measured with something else, before any test is run. Nothing in larsen-marx-mathematical-statistics or illowsky-introductory-statistics names it directly, though blocking and stratified sampling are the same underlying design move under different words. herzog-understanding-statistics's replication-crisis material documents the downstream symptom (published effects that do not replicate); this page names one mechanism that produces it.

Related

ramos-controlling-variables · mathematical-statistics · larsen-marx-mathematical-statistics · herzog-understanding-statistics · illowsky-introductory-statistics · mathematics