Confounding and Controls
A confounder is a hidden common cause that fakes a relationship; controls are the methods that strip its influence away.
What a Confounder Does
A confounder is a variable that influences both the supposed cause and the effect, creating an association between them that is not causal. Because it drives both, comparing groups that differ in the treatment also compares groups that differ in the confounder, and the two influences are tangled. Left uncontrolled, a confounder produces confident but false conclusions.
Controlling by Design
The strongest control is randomization: assigning the treatment by chance breaks any link between treatment and confounders, on average, including confounders no one thought of. This is why randomized experiments are the reference standard. When randomization is possible, it removes the need to enumerate every confounder in advance.
Controlling by Analysis
- Stratification: analyze within groups that share the confounder's value.
- Regression adjustment: include the confounder as a variable in the model.
- Matching: pair treated and untreated units with similar confounder profiles.
- Weighting: reweight the sample to balance the confounder across groups.
The Limit of Statistical Control
Analytical controls can only remove confounders that are measured. An unmeasured or unknown confounder passes straight through, and no amount of statistical sophistication recovers it. This is the core weakness of observational causal claims and the reason randomized evidence is prized: it controls the confounders you never listed.
Over-Control
Controlling for the wrong variable can also cause bias. Adjusting for a variable that lies on the causal path from cause to effect, a mediator, removes part of the very effect being measured, and adjusting for a common effect of two variables can create a spurious link. Deciding what to control for, and what to leave alone, requires a causal model, not a reflex to control for everything.