Confounder Adjustment

Confounder adjustment is how observational studies attempt to isolate the effect of an exposure from the influence of common causes. Design-stage approaches include restriction and matching; analysis-stage approaches include stratification, multivariable regression, propensity-score methods, and inverse-probability weighting. The goal is to compare exposed and unexposed groups that are alike with respect to the confounders, approximating the exchangeability a randomized trial achieves by design.

It is fundamental to causal inference in population science, where randomization is often impossible and confounding is the chief obstacle to credible effect estimates. Done well, adjustment lets a study report an effect that is "controlled for" known confounders.

The pitfalls are serious. Adjustment only handles measured confounders, so residual and unmeasured confounding can remain. Crucially, adjusting for the wrong variable backfires: conditioning on a mediator blocks part of the real effect, and conditioning on a collider opens new bias. This is why variable selection should follow a causal model (a DAG) rather than automated or significance-based procedures.

For example, to estimate whether a fitness program lowers cancer risk, a study might adjust for age, smoking, and income — but adjusting for "body weight after the program" would be a mistake, since weight may be on the causal pathway.