A topic in the Open Knowledge Graph — a free, open map of 15,290 topics and the order to learn them in.
Matching is a design strategy that pairs cases with controls on specific confounding variables (age, gender, etc.) to reduce confounding bias without necessarily losing statistical power. Matching can be 1:1, k:1, or frequency matching depending on study goals and resource constraints. Matched analyses require special statistical techniques such as conditional logistic regression to properly account for the matching structure and preserve bias reduction.
Compare unmatched and matched datasets for the same exposures and outcomes; visualize how matching reduces residual confounding.
Matching on a variable automatically controls for confounding without further adjustment. Overmatching on intermediate variables or strong correlates of exposure can unnecessarily decrease statistical efficiency and precision.
From your study of confounding, you know that a confounder is a variable associated with both the exposure and the outcome that can distort the apparent relationship between them. Matching is a design-level strategy to control confounding — rather than adjusting for imbalance after data collection (as regression and stratification do), matching prevents the imbalance from arising in the first place. In a case-control study, you have identified cases (people with the disease) and must select controls (people without it) for comparison. If you simply sample controls at random from the source population, they may differ from cases in age, sex, socioeconomic status, and dozens of other potential confounders. Matching selects controls who resemble cases on specified variables, so those variables cannot confound the exposure-disease comparison.
The mechanics are straightforward. For each case, you find one or more controls who share the value (or a close value) of the matched variable. 1:1 matching pairs each case with exactly one control — maximum comparability, moderate sample size. k:1 matching pairs each case with k controls — sacrifices some comparability for statistical power, and the efficiency gains from additional controls diminish past about 4:1. Frequency matching (or category matching) does not pair individuals but instead selects controls so that the distribution of the matched variable in the control group mirrors its distribution in the case group — easier to implement in large studies but less precise than individual matching.
The most important technical consequence of matched design is that matched data require matched analysis. This is the rule most commonly violated. When you match on age, you have deliberately removed the age variation that would otherwise confound — but you have also removed the age variation that your statistical model would use to estimate anything. An unmatched logistic regression applied to matched data ignores the pairing structure, treats the matched control as if it had been randomly selected, and produces biased and inefficient estimates. The correct method is conditional logistic regression, which conditions on matched sets rather than individuals, comparing the exposure status of a case to that of its matched controls. The odds ratio it produces properly accounts for the pairing.
A critically underappreciated failure mode is overmatching. If you match on a variable that is not truly a confounder — specifically, if you match on a variable that is a strong correlate of the exposure (an exposure proxy) or an intermediate step between exposure and outcome — you can inadvertently remove the variation in exposure needed to detect any association. Imagine studying smoking and lung cancer while matching on nicotine-stained fingers: you have now selected controls who are also heavy smokers, eliminating the exposure contrast between cases and controls. The result is a biased-toward-null estimate and wasted resources. The safeguard is to match only on variables that are independently associated with disease risk and are not caused by the exposure of interest.