A case-control study starts from the outcome and works backward, comparing people who already have a disease (the cases) with people who do not (the controls) to see whether their past exposure differs. Because it recruits on the basis of disease status rather than waiting for disease to appear, this observational study design answers questions about rare outcomes with a speed and economy that no forward-looking design can match. The trade-off is that it reports an odds ratio rather than a direct risk, and it is unusually sensitive to how the controls are chosen.
Imagine an outcome that affects one person in ten thousand. A cohort study would have to enroll and follow an enormous population for years to accumulate enough cases to analyze. A case-control study sidesteps that by going to where the cases already are, a clinic or a registry, and assembling a comparison group of people without the disease. In one efficient step it gathers enough cases to study a condition that a cohort study design could only reach at great cost. This is why case-control studies are the natural design for rare diseases, outbreak investigations, and conditions with long latency.
Cases and controls: the two decisions that matter most
A case-control study is only as good as its definitions.
- Case definition. Cases should be identified by explicit, consistently applied criteria, ideally incident (newly diagnosed) cases rather than prevalent ones, so the study reflects causes of disease onset rather than causes of survival.
- Control selection. Controls must come from the same source population that produced the cases and must represent the exposure distribution of that population. Choosing controls who differ systematically from the source population is the single most common way a case-control study goes wrong.
Get these two right and the design is powerful. Get control selection wrong and no amount of analysis will rescue the result.
The odds ratio, and why this design reports it
Because a case-control study begins with a fixed group of cases and controls rather than a natural population, it cannot measure incidence, so it cannot report a relative risk directly. What it can estimate is the odds ratio: the odds of exposure among cases divided by the odds of exposure among controls. For a rare outcome the odds ratio closely approximates the relative risk, which is what makes the design interpretable. When you have your two-by-two counts, our odds ratio calculator returns the estimate and its confidence interval. Reading that number correctly, and knowing when it does and does not approximate risk, is essential to reporting a case-control study honestly.
Matching, and its hidden cost
Researchers often match controls to cases on variables such as age and sex to remove those as confounders. Matching can improve efficiency, but it is not free: a matched design requires a matched analysis, such as conditional logistic regression, and you can no longer study the matched variable as a risk factor because you have forced it to be equal across groups. Over-matching, matching on a variable on the causal pathway between exposure and outcome, can even bias the result toward the null. Match deliberately and analyze accordingly, or you will undo the benefit.
Two biases are intrinsic to looking backward:
- Recall bias. Cases, motivated by their diagnosis, may remember and report past exposures differently from controls. Because exposure is measured after the outcome, this is a structural risk, not an oversight.
- Selection bias. If the route by which cases and controls entered the study is related to exposure, the odds ratio is distorted before any analysis begins.
Structured appraisal helps reviewers weigh these threats; the quality appraisal with the Newcastle-Ottawa Scale was designed for exactly the case-control and cohort domains of selection, comparability, and exposure ascertainment.
Nested case-control and case-cohort variants
When a cohort already exists, a nested case-control study draws cases and controls from within it, combining the clean exposure measurement of a cohort with the analytic efficiency of a case-control comparison. These hybrid designs are increasingly common with large biobanks and electronic health records, and they blunt recall bias because exposure was recorded before anyone became a case.
The textbook line that the odds ratio approximates the risk ratio only when the disease is rare is true for one specific way of choosing controls, and obscures a more useful fact: with the right sampling, the odds ratio estimates a real effect measure with no rarity assumption at all. Three control-sampling schemes are worth knowing by name. Cumulative (case-base) sampling takes controls from those still disease-free at the end of follow-up, and here the odds ratio approximates the risk ratio only when the outcome is rare. Density (risk-set) sampling, where each case is matched to controls sampled from those still at risk at the moment the case occurs, makes the odds ratio estimate the incidence rate ratio directly, rare disease or not. Case-cohort sampling draws controls from a random subcohort selected at baseline and lets the odds ratio estimate the risk ratio. The practical lesson is to decide the sampling scheme deliberately and then interpret the odds ratio as the measure that scheme actually targets, rather than reciting the rare-disease caveat by reflex.
A crude odds ratio from the full two-by-two table assumes the groups differ only in exposure, which they rarely do. The classical adjustment is the Mantel-Haenszel odds ratio, which pools the exposure-disease association across strata of a confounder and is still the transparent way to show whether adjustment moves the estimate. When confounders are numerous or continuous, unconditional logistic regression is the standard tool for an unmatched design, estimating an adjusted odds ratio for every covariate at once. A matched design changes the analysis, not just the recruitment: the matched sets must be kept together with conditional logistic regression, because an ordinary model that ignores the matching is biased. Watch for the same overadjustment trap that matching creates, never adjust for a variable on the causal pathway from exposure to outcome (a mediator) or for a collider, since conditioning on a common effect of exposure and outcome opens a spurious association rather than closing a real one.
Because exposure is reconstructed after the outcome, measurement error is a structural feature of the design, and its direction is partly predictable. Non-differential misclassification, error that is equally likely in cases and controls, generally biases the odds ratio toward the null, so a real effect is underestimated and a study that found nothing may simply have measured exposure badly. Differential misclassification, the recall bias where cases scrutinise their past more than controls do, can bias the estimate in either direction and is the more dangerous because it can manufacture an association that is not there. The mature response is a quantitative bias analysis: rather than listing recall bias as a limitation, specify plausible sensitivity and specificity of exposure measurement and recompute the odds ratio under those assumptions, or report an E-value stating how strong an unmeasured confounder would need to be to explain the result away.
library(epitools)
library(survival)
# Crude odds ratio with a confidence interval from the two-by-two table
oddsratio(table(exposure, disease))
# Adjust for a confounder by stratification (Mantel-Haenszel)
mantelhaen.test(table(exposure, disease, stratum))
# Unmatched design: adjusted odds ratios from logistic regression
fit <- glm(disease ~ exposure + age + sex, data = d, family = binomial)
exp(cbind(OR = coef(fit), confint(fit)))
# Matched design: conditional logistic regression keeps the matched sets together
clogit(disease ~ exposure + strata(matched_set), data = d)
The decision rule is simple to state. If the outcome is rare or slow to develop, a case-control study reaches an answer efficiently. If the exposure is rare instead, a cohort is better. If you only need to know how common something is right now, a cross-sectional snapshot is fastest. Once the design is settled, selecting an analysis that respects it, conditional logistic regression for matched data, ordinary logistic regression otherwise, is what turns a sound design into a defensible result.
