Preserve matched sets in case-control analysis

Recover the sampling design

In a matched case-control study, the analysis must reflect how controls were selected. Start with the matching protocol, not the name of a statistical test. Record the matching factors, the permitted number of controls per case and whether a person can appear in more than one risk set. Distinguish individual matching from frequency matching.

The survival package conditional-logistic documentation shows how matched-set strata enter the conditional likelihood. A matched design is not represented adequately by deleting the set identifier and fitting an ordinary exposure-only logistic model without a design-based rationale.

Audit the analysis table

Check that each matched-set identifier contains the expected case and control records. Verify the case indicator, exposure timing and matching-variable definitions. Investigate duplicated rows, sets with no case and sets changed by missing covariates. Document whether exclusions remove individuals or whole sets and how that affects the target comparison.

Inspect exposure variation within sets. Sets in which everyone has the same exposure provide no information about that exposure's conditional contrast. A large total row count can therefore conceal limited information for the coefficient of interest. Covariates constant within every set cannot have their separate effects estimated by the conditional likelihood.

Hypothetical example

In a fictional one-to-one matched study, 12 pairs have an exposed case and unexposed control, while four have an unexposed case and exposed control. The simple matched-pair exposure odds ratio is 12/4 = 3. Pairs concordant for exposure do not enter that ratio.

This calculation is for the simple paired exposure comparison. It is not an adjusted estimate for additional covariates, and its interpretation depends on the sampling design. It does not turn the odds ratio into an absolute risk or risk difference.

Fit and report the intended model

Specify the matched-set strata, covariates, missing-data handling and likelihood or approximation used. Check convergence, sparse patterns and uncertainty. If weights are needed, confirm that the chosen implementation supports their intended meaning; software options should not be assumed interchangeable.

Report the number of cases, controls and contributing sets, along with the estimate and confidence interval. Explain the control-sampling method so readers can interpret the measure correctly. Use the analysis test-oracle worksheet to retain the simple 12-to-four synthetic check alongside the production model tests.

Source verification: 6 November 2026. Examples are hypothetical.