Study Matching and Stratification

Matching and stratification are design devices used to control confounding by building balance for known factors into a study from the outset. Matching pairs or groups subjects so that comparison groups share the same distribution of a confounder, while stratification divides subjects into homogeneous strata within which comparisons are made. Both are ways of making comparison groups more alike on selected variables so that the contrast of interest is less distorted by those variables.

Definition

Matching is a design technique that selects comparison subjects to share the distribution of one or more confounders with the index subjects, and stratification is the division of subjects into subgroups (strata) defined by confounders so that exposure-outcome comparisons are made within homogeneous strata.

Scope

The entry covers the rationale for matching and stratification, the difference between individual and frequency matching, the use of strata in both observational studies and randomized trials, and the analytic implications (such as the need for matched or stratified analysis). It is framed as a methodological reference on confounding control by design and does not provide clinical instructions.

Key concepts

• Confounding control by design
• Individual (pair) matching versus frequency matching
• Strata and within-stratum comparison
• Stratified randomization in trials
• Matched analysis (conditional methods)
• Overmatching
• Loss of efficiency from matching on non-confounders

Mechanisms

Both techniques remove or reduce confounding by a chosen variable before analysis. Matching forces the matched factor to have the same distribution in the groups being compared, so it can no longer confound the association, but it requires an analysis that respects the matched structure; analyzing matched data as if unmatched can bias results. Stratification partitions subjects into strata within which the confounder is essentially constant, estimates the association within each stratum, and combines the stratum-specific estimates. In randomized trials, stratified randomization performs the allocation separately within strata to keep important prognostic factors balanced across arms, usually combined with blocking.

Clinical relevance

Recognizing whether a study controlled confounding by matching or stratification, and whether it analyzed the data accordingly, is part of appraising whether an observed association is credible. This entry describes design and analysis methodology for research and is not a source of diagnostic or treatment guidance.

Evidence & guidelines

Methodological literature distinguishes the design act of matching from the analytic act of stratified or matched analysis, and emphasizes that matched designs require matched analyses to avoid bias. Guidance on stratified randomization in trials notes that it is most useful in smaller studies and should be paired with blocking, and standard epidemiology texts set out when matching improves efficiency and when overmatching on a non-confounder harms it.

History

Matching has long been used in case-control studies of chronic disease to control strong confounders such as age and sex, and Breslow and Day's 1980 monograph codified the conditional (matched) analysis these designs require. Stratified analysis traces to the Mantel-Haenszel methods of the mid-twentieth century, and stratified randomization was adopted in clinical trials to keep prognostic factors balanced across treatment arms, with later methodological reviews clarifying when it adds value.

Debates

When does matching help, and when does it backfire?

Matching on a genuine confounder can improve efficiency, but matching on a variable that is not a confounder, or that lies on the causal pathway, can reduce efficiency or introduce bias (overmatching); the decision depends on the causal structure, not convenience.

Is stratified randomization necessary in large trials?

Stratification keeps key prognostic factors balanced and is most valuable in smaller trials, while in large trials simple randomization tends to balance factors on its own; over-stratification can create many sparse strata and complicate the design.

Key figures

• Norman Breslow
• Nicholas Day
• Kenneth Rothman
• Sander Greenland
• Neil Pearce

Seminal works

• breslow-day-1980-matching
• pearce-2016-matched
• kernan-1999-stratified

Frequently asked questions

What is the difference between matching and stratification?

Matching is a sampling decision made when subjects are selected (choosing comparison subjects to share a confounder's distribution), while stratification divides subjects into subgroups defined by a confounder and compares exposure and outcome within those subgroups; matched data also require a matched analysis.

What is overmatching?

Overmatching is matching on a variable that should not be matched, such as one that is not a confounder or that lies on the causal pathway between exposure and outcome; it can reduce statistical efficiency or bias the estimate rather than improve control of confounding.