Matching in Case-Control Studies

Overview

Matching deliberately selects controls so that the distribution of one or more characteristics resembles that of cases. In individual matching, each case is linked to one or more controls with specified values or ranges. In frequency matching, controls are selected so that the overall distribution of matching factors approximates the case distribution.[1–3]

Matching can improve efficiency, ensure representation across important strata and manage strong confounding by variables such as age or calendar time. It also changes the sampling design. A matched factor is constrained by selection and must be handled appropriately in analysis. Matching does not automatically remove confounding, and careless matching can reduce validity, create recruitment difficulty or prevent estimation of the matched variable’s association with outcome.[2–4]

This lecture develops the design, conduct and analysis logic of matching in case-control studies.

Why this matters

Matching is often added because it appears to make groups “comparable”. That goal is too vague. A protocol should explain which source of confounding or inefficiency is addressed and why matching is preferable to unrestricted sampling with analytical adjustment.

Poor matching can cause:

• overmatching, when controls are made similar on a consequence or close correlate of exposure;
• inability to study the effect of a matched factor;
• exclusion of cases for whom no eligible control is found;
• very narrow calipers that make recruitment slow and unrepresentative;
• selection of friend or family controls with shared exposures;
• unmatched analysis that discards the design; or
• residual confounding when broad matching categories are treated as exact control.

Key concepts

Individual matching

Each case defines a matched set. Controls satisfy prespecified criteria, such as exact sex, age within a caliper, district and index date. The set identifier must remain in the dataset. Analysis compares exposure discordance within sets rather than treating all participants as independent.

Individual matching may be one-to-one, one-to-many or variable-ratio. More controls per case can improve precision when controls are inexpensive, but the gain diminishes as the ratio increases. A flexible ratio can retain cases when only one or two controls are available, provided the analysis accommodates variable set sizes.[1]

Frequency matching

Controls are sampled to reproduce the aggregate distribution of matching factors among cases. For example, if 30% of cases are aged 30–39 years, 45% are 40–49 and 25% are 50–54, the control sample may be selected in similar proportions. There is no individual case-control link.

Frequency matching generally uses an unmatched analytical model with the matching variables included. Failing to adjust for frequency-matching factors can introduce confounding because control selection was conditioned on those variables.[2]

Matching factors

A defensible factor usually:

• is a strong determinant of the outcome;
• may be associated with the exposure;
• is measured accurately before selection;
• is not caused by the exposure;
• is not an intermediate variable; and
• has enough available controls within strata.

Causal diagrams help distinguish confounders from mediators, colliders and exposure proxies. Matching on a collider can create association between exposure and unmeasured causes of selection. Matching on a strong exposure correlate that is not a confounder may reduce exposure contrast and efficiency.[3,4]

When matching is useful

DECISION GUIDE

Questions before matching

• Is the proposed factor a strong cause of the outcome and a plausible confounder?
• Would unrestricted sampling leave sparse or highly imbalanced strata?
• Is the factor measured before control selection with little error?
• Could matching prevent estimation of an effect that matters scientifically?
• Are enough controls available within each matching stratum?
• Can matched-set identifiers and factors be retained through data management and analysis?
• Would stratification or regression adjustment be simpler and equally effective?

Matching is especially useful for tightly controlling age, sex, site, calendar time or risk-set membership when these variables are central to the sampling design. It may be unnecessary when the source population is large, factors are readily modelled and matching would complicate recruitment.

Research question and methodological fit

The research question remains whether previous gestational diabetes is associated with first myocardial infarction. Age and district are plausible common causes of exposure documentation and outcome. Individual matching can ensure every case has controls from comparable age and geographic strata.

The design should define whether the estimand is conditional on the matched factors or intended to generalise across their distribution. It should also clarify that the association of age or district with outcome cannot be estimated in the ordinary way after exact individual matching because their distributions are fixed within sets.

Design, setting and study population

DESIGN PROFILE

Individually matched case-control study

Core structure: Each case is linked to one or more controls selected from the same study base according to prespecified matching factors and a comparable index time.

Best suited for: Efficient control of strong design variables, risk-set sampling and assurance that important strata contribute controls.

Important limitations: Recruitment can fail in sparse strata; matched factors require design-aware analysis; overmatching can reduce exposure information or introduce bias.

The protocol must define the matching algorithm before recruitment. “Similar age” is not reproducible. A specification might be exact district, age within two completed years and at-risk status on the case index date. If several controls qualify, one or more should be selected randomly rather than by convenience.

Pair, set and variable-ratio matching

One-to-one matching creates pairs. One-to-two or one-to-four designs create sets. Variable-ratio matching allows each case to have a different number of controls and may improve feasibility. The maximum ratio, minimum accepted controls and action when no control is found should be prespecified.

A case should not be discarded automatically because one planned control declines. If at least one valid control remains, the set may still contribute to conditional analysis. If no control is recruited, the case contributes no within-set exposure contrast and ordinarily cannot enter the primary matched analysis.

Sampling and recruitment

PROTOCOL CHECKLIST

Matching algorithm to prespecify

• Identify the matching factors and causal rationale for each.
• Define exact matches, categories or numerical calipers.
• Specify the case index date and control eligibility at that date.
• Define the number or range of controls per case.
• State how multiple eligible controls are sampled.
• Specify whether a person may serve as a control for more than one risk set.
• Define procedures for refusal, unavailable records and unmatched cases.
• Preserve the matched-set identifier and factor values in the analytical dataset.

Caliper width

A very narrow age caliper improves similarity but may exclude cases and extend recruitment. A broad caliper improves feasibility but leaves residual age differences. The trade-off should be evaluated with expected case distributions and frame counts before the study begins.

Matching with replacement and repeated selection

In risk-set sampling, a person may be eligible as a control for several case times and can later become a case. Repeated selection reflects repeated person-time sampling and may be valid. The protocol and statistical team should agree how repeated identifiers are represented and whether robust variance or specialised methods are needed.

Variables, constructs and measurements

Illustrative matched-set data structure

illustrative

Table 1. One case with two individually matched controls

set_id	role	age at index	district	previous gestational diabetes	index date
001	Case	47	A	Yes	2026-03-14
001	Control	46	A	No	2026-03-14
001	Control	48	A	Yes	2026-03-14

Note: The set_id is essential. Exposure information is illustrative.

The dataset should also contain continuous values of matching variables even when categories were used. This permits description of residual differences and sensitivity analyses. Do not remove the matching variables because “they are already controlled”.

Procedures and data collection

Matching should occur from eligibility variables, not from exposure information. Recruiters should not choose a control because that person appears unexposed. Automated candidate lists with random ordering reduce discretion.

Quality monitoring should track:

• cases awaiting controls;
• number of eligible candidates per set;
• refusal and missing-record rates by matching stratum;
• time from case identification to completed set;
• distribution of caliper differences; and
• frequency of protocol deviations.

If certain districts or age groups repeatedly lack controls, the investigators should reconsider the frame or matching specification rather than silently widening criteria case by case.

Bias, validity and quality assurance

Matching does not control all confounding

Only the matched factors are constrained, and even they may retain residual confounding if categories are broad or measurement is inaccurate. Other common causes require measurement and adjustment.

Matching can induce confounding in the sampled data

Because control selection depends on matching factors, the relationship between exposure and the matching factor among sampled controls may differ from that in the source population. Analysis must condition or adjust for matching variables to recover the intended comparison.[2]

Broken matched sets

If data are missing for one member, the analytical consequences depend on role and method. A set with a case but no valid control cannot contribute to conditional logistic regression. A set with one remaining control may contribute even if two were planned. Imputing the missing exposure should respect the clustered set structure and uncertainty.

APPRAISAL CHECKLIST

Appraising a matched case-control study

• Is there a clear causal or efficiency rationale for every matching factor?
• Are factors measured before selection and unlikely to be mediators or colliders?
• Are calipers and matching ratios explicit and feasible?
• Were eligible controls sampled without knowledge of exposure?
• Are unmatched cases, refusals and incomplete sets reported?
• Does the analysis preserve individual sets or adjust for frequency-matching factors?
• Are residual confounding and overmatching considered in interpretation?

Ethical and governance considerations

Interface with statistical analysis

For individually matched studies, conditional logistic regression estimates an odds ratio by comparing exposure within matched sets. Sets with no exposure variation provide no information about the exposure coefficient. Crude unmatched tables can be shown descriptively but should not replace matched analysis.[1,5]

For frequency matching, ordinary logistic regression is generally used with matching factors included as covariates. Additional confounders can be added based on the causal model. Interaction terms may assess whether associations differ across prespecified strata, but matching itself is not effect modification.

Conditional versus unconditional analysisMMed extension

Conditional logistic regression directly respects individual matching and is the usual primary method. Unconditional logistic regression that includes detailed matching factors can sometimes be valid and more efficient, particularly with many controls or sparse sets, but it requires correct modelling of the factors and should be justified prospectively. Simply ignoring the set structure is not acceptable.[1,5]

Practical decisions when matching fails

A matching plan should anticipate that some cases will have few eligible controls. The study team can examine frame counts by age, site and calendar period before recruitment, simulate expected set completion, and prespecify a hierarchy of permissible actions. One defensible hierarchy is to retain the original factors, accept a variable number of controls, and widen a numerical caliper only through a formally approved protocol amendment applied prospectively. Quietly changing criteria for selected cases makes inclusion depend on investigator discretion.

When a case remains unmatched, report the case and compare its measured characteristics with matched cases. Excluding unmatched cases changes the case series and may narrow generalisability, especially when matching failure is concentrated in old age, remote districts or unusual calendar periods. It is not valid to add an unmatched convenience control and then label the pair matched.

The same discipline applies after data collection. If a control lacks the primary exposure, the set may retain other valid controls. If the case lacks exposure, the set cannot inform the primary contrast. Sensitivity analyses can compare complete matched sets with carefully justified imputation, but the set structure and selection process must remain visible.

Matching ratio and information

Adding controls is most useful when cases are scarce and controls are inexpensive. The gain depends on exposure frequency and within-set discordance, not only the numerical ratio. A fourth nearly identical control may add little information, whereas improving exposure measurement for existing set members may be more valuable. Sample-size planning should therefore incorporate the planned matching ratio and expected correlation induced by matching rather than applying an unmatched formula.

Worked example

WORKED EXAMPLE

Designing age-and-district matched sets

The investigators choose individual matching on district and age within two years. Body mass index is not matched because it may partly follow previous gestational diabetes and is of scientific interest. Current healthcare utilisation is not matched because it may be influenced by cardiovascular symptoms and record access.

For each incident case, the population register generates all eligible women who are at risk on the index date. Two controls are randomly selected. When one declines, another candidate is randomly selected from the same list and the refusal remains documented. If only one valid control can be recruited, the set is retained.

The dataset stores set_id, exact age, district and index date. Conditional logistic regression estimates the association of previous gestational diabetes with myocardial infarction while additional prespecified confounders are included. The Methods section states the matching factors and analytical method.

Suggested protocol or reporting wording

Common mistakes

Knowledge check

Does individual matching on age allow the study to estimate the effect of age on outcome?

Answer and reasoning

Not in the ordinary matched analysis. Age is constrained within sets and cases are not compared with controls across the full age distribution. The design uses age to control sampling, not to estimate its independent association.

Why must frequency-matching variables usually be included in the regression model?

Answer and reasoning

Control sampling was conditioned on their distribution. Adjustment restores the intended comparison and addresses confounding by those factors. Frequency matching does not create individual pairs that can be ignored.

A case has one valid control although two were planned. Must the entire set be discarded?

Answer and reasoning

Usually no. A case with at least one valid matched control can contribute to conditional analysis. The actual ratio should be retained and reported; the software can accommodate variable set sizes.

Why might matching on current healthcare utilisation bias the illustrative study?

Answer and reasoning

Healthcare utilisation may be affected by early cardiovascular symptoms or by consequences of previous gestational diabetes. Conditioning control selection on it could block part of a pathway or create selection through common causes.

What information makes matched analysis possible?

Answer and reasoning

The matched-set identifier, role, matching-factor values and index date must be retained. Without the set identifier, the intended within-set comparison cannot be reconstructed reliably.

Key take-home messages

SUMMARY

Key take-home messages

• Matching deliberately changes control sampling and must have a stated rationale.
• Individual matching creates sets; frequency matching balances aggregate distributions.
• Matching does not by itself remove confounding.
• Avoid matching on mediators, colliders or unnecessary exposure proxies.
• Calipers, ratios, candidate selection and failure rules should be prespecified.
• Individual matching ordinarily requires conditional analysis; frequency matching requires adjustment.

Reporting should describe matching criteria, numbers of controls per case, unmatched cases and the matched analysis in accordance with STROBE.[6]

References

1. Keogh RH, Cox DR. Case-control studies. Cambridge: Cambridge University Press; 2014. https://doi.org/10.1017/CBO9781139094757
2. Pearce N. Analysis of matched case-control studies. BMJ. 2016;352:i969. https://doi.org/10.1136/bmj.i969
3. Mansournia MA, Hernán MA, Greenland S. Matched designs and causal diagrams. Int J Epidemiol. 2013;42(3):860–869. https://doi.org/10.1093/ije/dyt083
4. Rothman KJ, Greenland S, Lash TL. Modern epidemiology. 3rd ed. Philadelphia: Lippincott Williams & Wilkins; 2008.
5. Breslow NE, Day NE. Statistical methods in cancer research. Volume I: The analysis of case-control studies. Lyon: International Agency for Research on Cancer; 1980.
6. von Elm E, Altman DG, Egger M, Pocock SJ, Gøtzsche PC, Vandenbroucke JP; STROBE Initiative. The Strengthening the Reporting of Observational Studies in Epidemiology statement. PLoS Med. 2007;4(10):e296. https://doi.org/10.1371/journal.pmed.0040296