Matched Case-Control Study
- Introduced by
- Brian MacMahon and others; systematised by Schlesselman (1982)
- Year
- 1950s–1970s
- Type
- Observational analytic design
- Family
- Clinical / epidemiology
What does it do?
A matched case-control study is an observational epidemiological design in which each case (a person with the disease or outcome of interest) is paired with one or more controls (persons without the outcome) who share one or more characteristics — such as age, sex, or clinical setting — to control confounding.
Exposure history is then compared between cases and their matched controls to estimate the odds ratio of the exposure-disease association.
The intuition behind the method
Think of it like a detective pairing suspects with look-alikes. If you want to know whether a particular exposure caused a disease, you find a patient with the disease and then locate one or more similar people — same age, same sex, same hospital — who did not get the disease. By keeping those background features identical, any difference in exposure history between the pair is far more likely to reflect a real link rather than a confounding factor. Matching tightens the comparison so the signal is cleaner, at the cost of some flexibility in the analysis.
History
The case-control design was formalised in epidemiology in the 1950s, with landmark studies by Doll and Hill (1950) on smoking and lung cancer using hospital-based controls. The principle of matching to improve comparability was applied early and codified by Brian MacMahon and colleagues. James Schlesselman's 1982 textbook Case-Control Studies established the analytical framework for matched designs, including the matched odds ratio and conditional logistic regression. Nathan Mantel and William Haenszel's 1959 paper provided foundational stratified methods that underpin matched analysis, and the conditional logistic regression approach was developed through the 1970s–1980s, becoming the standard analytical tool for individually matched case-control data.
When to use it
Use a matched case-control study when studying a rare disease or outcome where a full cohort study would require impractically large samples or very long follow-up; when specific confounders (such as age, sex, or clinical centre) need tight control; and when resources permit careful individual matching and conditional analysis. It is especially valuable in hospital-based settings and outbreak investigations. Do not use matching when the matching variables are potential intermediates on the causal pathway (over-matching), when the number of cases is very small and matching partners cannot be found, or when time-varying exposures make static matching inappropriate — in those situations a case-crossover or cohort design may be preferable. Matching does not eliminate unmeasured confounding.
Limitations and considerations
- Cannot estimate the prevalence or incidence of the outcome — only the odds ratio for the exposure-outcome association.
- Matching variables cannot themselves be studied as exposures unless complex analytic methods are used.
- Over-matching (matching on a variable associated with exposure but not disease) inflates the required sample size and biases the odds ratio toward the null.
- Finding adequate matched controls can be difficult, particularly for rare combinations of matching criteria, leading to unmatched cases that must be excluded.
- The matched structure must be maintained in analysis; ignoring it (using unconditional logistic regression) produces biased estimates.
Source conditions do not replace independent scientific review or validation by an analysis engine.
Strengths
- Efficiently controls for strong confounders (age, sex, clinical centre) by design rather than solely by statistical adjustment.
- Well-suited to rare outcomes where a full cohort is not feasible; a fraction of the unmatched sample can achieve comparable precision.
- Reduces variability within matched sets, often increasing statistical efficiency for a given sample size.
- Widely accepted and well-understood in clinical and epidemiological journals; conditional logistic regression is standard and implemented in all major statistical packages.
Applications
- Investigating risk factors for rare cancers by matching hospital patients with cancer to controls from the same hospital on age, sex, and admission ward.
- Outbreak investigations in which cases of foodborne illness are matched to neighbourhood controls on geographic location and meal occasion.
- Pharmacoepidemiological studies examining adverse drug reactions, with cases matched to drug-exposed controls on indication, age, and comorbidity.
- Occupational epidemiology studies of rare exposures (e.g., specific chemical agents) with industry-matched controls to remove occupational confounding.
- Genetic epidemiology where cases and controls are ethnicity-matched to minimise population stratification before genotyping.
- Emergency and surgical studies where hospital admission provides a natural matching frame for acute exposure–outcome questions.
How it works
How it works
Formulate a precise case definition specifying the disease or outcome, its diagnostic criteria, the eligible population (source population), and the time window. Identify the exposure of interest and the confounders you intend to control by matching. Pre-specify the matching ratio (1:1, 1:2, etc.) and the matching variables.
Recruit incident or prevalent cases from a defined source such as hospital records, disease registries, or a surveillance system. Confirm each case meets the case definition. Select controls from the same source population — people who would have been eligible to become a case — using hospital-based, population-based, or density sampling strategies.
For each case, select one or more controls who share the values of the matching variables (e.g., same five-year age band and sex). Individual (one-to-one or one-to-many) matching ties each control to a specific case; frequency matching balances distributions across the study as a whole. Document the matching ratio and any caliper widths used.
Ascertain exposure status for both cases and their matched controls, using the same measurement method for both groups to avoid differential misclassification. Common sources include interviews, medical records, biobank data, or administrative databases. Blinding outcome assessors to case/control status reduces information bias.
Analyse matched data with conditional logistic regression (or McNemar's test for 1:1 matched pairs) to preserve the matched structure. For a simple 1:1 matched design the matched odds ratio is estimated as OR = b/c, where b = discordant pairs with exposed case and unexposed control, and c = discordant pairs with unexposed case and exposed control. Adjust for additional covariates within the conditional model.
Formula
OR = (b / c)
Express results as matched odds ratios with 95% confidence intervals. Evaluate whether the matching variables were appropriate, check for over-matching (matching on an exposure intermediate), and assess residual confounding. Discuss external validity with respect to the source population.
Common pitfalls
- Over-matching: selecting matching variables that lie on the causal pathway between exposure and disease (intermediates) removes the variation needed to detect the association and biases OR toward 1.0.
- Breaking the match in analysis: analysing matched data with standard (unconditional) logistic regression discards the efficiency gains of matching and can introduce bias — always use conditional logistic regression or McNemar's test.
- Recall bias: if cases recall exposures more vividly than controls (differential recall), the odds ratio will be overestimated; use objective records or biomarkers where possible.
- Ignoring unmatched cases: excluding cases for whom no matched control was found can introduce selection bias if excludability is related to exposure.
- Insufficient control for residual confounding: matching on broad categories (e.g., ten-year age bands) leaves within-category confounding unaddressed; refine categories or adjust for age continuously in the model.
Frequently asked questions
Why must I use conditional logistic regression rather than ordinary logistic regression?
In an individually matched study, the case and its matched controls form a stratum. Unconditional logistic regression treats all strata as a single pooled sample, which produces biased estimates because the matched-set intercepts are treated as fixed parameters — leading to the incidental parameters problem. Conditional logistic regression conditions on the sufficient statistic for each matched set, eliminating the stratum-specific intercepts and yielding an unbiased matched odds ratio.
What is over-matching and why is it harmful?
Over-matching occurs when you match on a variable that is associated with the exposure but not independently with the disease (or is an intermediate on the causal path). Because cases and controls then have similar exposure histories by design, the contrast between groups is reduced, the odds ratio is biased toward 1, and a larger sample is needed to achieve the same power. Avoid matching on variables that could be downstream consequences of the exposure.
How do I choose the matching ratio?
A 1:1 ratio is the most common and easiest to manage. Increasing to 1:2 or 1:3 controls per case improves precision when cases are rare but controls are plentiful; the gain in efficiency diminishes beyond 1:4 and rarely justifies the added complexity. If cases are very rare and controls abundant, a 1:4 ratio is often considered the practical maximum.
Can I study multiple exposures in the same matched case-control study?
Yes. The conditional logistic regression model can include multiple exposure variables simultaneously, allowing assessment of several risk factors while preserving the matched structure. However, each additional exposure adds parameters and requires an adequately sized sample; avoid data dredging by pre-specifying primary and secondary exposures in a study protocol.
What is the difference between individual matching and frequency matching?
Individual (pair) matching ties each specific case to one or more specific controls on matching variables, requiring conditional analysis. Frequency matching ensures that the overall distribution of matching variables is balanced between cases and controls but does not create case-specific pairs; it can be analysed with unconditional logistic regression adjusting for the matching variables. Individual matching provides tighter control but is logistically more demanding and constrains the analysis.
Related methods
Sources
Rothman, K. J., Greenland, S., & Lash, T. L. (2008). Modern Epidemiology (3rd ed.). Lippincott Williams & Wilkins.
ISBN: 978-0781755474Copied.Schlesselman, J. J. (1982). Case-Control Studies: Design, Conduct, Analysis. Oxford University Press.
ISBN: 978-0195029338Copied.