Case–Control Study
A case–control study is an observational epidemiological study design that samples participants based on disease outcome status—comparing incident cases against non-diseased controls—and retrospectively ascertains prior exposures to estimate the odds ratio.
Case–control study is an analytical observational study design in epidemiology and clinical biostatistics in which investigators sample subjects according to their disease status—enrolling individuals who have developed a specific outcome of interest (cases) alongside a comparable reference sample of individuals free of that outcome (controls)—and then look backward in time to compare prior exposure histories between the two groups. Pioneered in the landmark 1950 investigations of cigarette smoking and bronchogenic carcinoma by Richard Doll and Austin Bradford Hill in the United Kingdom and Ernst Wynder and Evarts Graham in the United States, and later placed on a rigorous statistical foundation by Jerome Cornfield (1951), Nathan Mantel and William Haenszel (1959), and Norman Breslow and Nicholas Day (1980), the case–control architecture solves a major logistical limitation of the prospective cohort study: the prohibitive sample sizes and multi-decade follow-up periods required to study rare diseases or outcomes with long induction and latency periods.
Study Architecture and the 2 × 2 Contingency Table
Outcome-Selective Sampling Strategy
Whereas a cohort study defines groups by exposure status at baseline and follows them forward to measure disease incidence, a case–control study reverses the sampling direction by selecting all (or a random fraction f1) of incident cases arising within a defined source population over a specified risk period, together with a much smaller sampling fraction (f0 ≪ f1) of unaffected controls drawn from the very same source population. Subjects in both groups are cross-classified by prior exposure status in a standard 2 × 2 contingency table:
| Prior Exposure Status | Cases (Diseased, D+) | Controls (Non-Diseased, D−) | Row Totals |
|---|---|---|---|
| Exposed (E+) | a | b | a + b |
| Unexposed (E−) | c | d | c + d |
| Column Totals | a + c (N1) | b + d (N0) | N = N1 + N0 |
The Exposure Odds Ratio (OR) and Cornfield's Theorem
Because the investigator fixes the ratio of cases (N1) to controls (N0) by design (for example, 1:1 or 1:4 matching), a case–control sample cannot directly compute population disease risk a / (a + b) without external knowledge of the population sampling fractions. Instead, the primary measure of association is the odds ratio (OR)—the ratio of the odds of prior exposure among cases (a / c) to the odds of prior exposure among controls (b / d):
In 1951, Jerome Cornfield proved the symmetry property of the odds ratio—showing that the exposure odds ratio equals the disease odds ratio—and demonstrated that under the rare-disease assumption (when cumulative disease incidence is low, typically < 5%–10% in both exposed and unexposed strata so a ≪ b and c ≪ d), the odds ratio closely approximates the risk ratio (relative risk, RR) that would be obtained from a full cohort study:
On the natural logarithmic scale, the asymptotic standard error of ln(OR) is estimated by Woolf's formula:
yielding the 95% confidence interval OR × exp( ±1.96 × SE[ln(OR)] ). When adjusting for multiple continuous and categorical confounders, unconditional logistic regression estimates adjusted log-odds ratios βk, while matched sets are analyzed using Mantel–Haenszel stratification or conditional logistic regression (analogous to stratified partial likelihood in the proportional hazards model).
Control Sampling Schemes and What the Odds Ratio Estimates
Modern epidemiological theory (Miettinen, 1976; Rothman, Greenland, & Lash, 2008) emphasizes that the rare-disease assumption is not required when controls are sampled dynamically from the person-time at risk:
| Control Sampling Design | Timing of Control Selection | Underlying Source Population Parameter Estimated by OR | Need for Rare-Disease Assumption? |
|---|---|---|---|
| Incidence-Density (Risk-Set) Sampling | Controls are sampled longitudinally at the calendar times (ti) when each incident case occurs (nested within risk sets). | Incidence Rate Ratio (IRR) (direct ratio of hazard/incidence rates λ1 / λ0) | No — valid regardless of how common the disease is. |
| Case–Cohort (Case–Base) Sampling | Controls are selected as a random subcohort of the total baseline population at the start of follow-up (t = 0), regardless of future disease status. | Cumulative Incidence Ratio (Risk Ratio, RR) over the follow-up period | No — directly estimates RR without rarity assumptions. |
| Cumulative (Exclusive / Survivor) Sampling | Controls are sampled at the end of the follow-up period from individuals who remained disease-free throughout the study. | Disease Odds Ratio, which approximates RR or IRR when disease risk is low | Yes — requires cumulative incidence < 10% to approximate RR. |
Sources of Bias, Confounding, and Modern Variants
Selection Bias (Berkson's Fallacy) and Recall Bias
Because both exposure and outcome have already occurred when subjects are enrolled, case–control studies require careful protection against two systematic biases prior to evaluating causal inference under the Bradford Hill criteria:
- Selection Bias and Berkson's Bias: Controls must represent the exposure distribution of the source population that generated the cases. When both cases and controls are sampled from hospitalized inpatients (Berkson's fallacy, 1946), differing hospitalization rates for the exposure and the disease distort the observed odds ratio unless control diagnoses are independent of the exposure under study. In infectious disease surveillance, the test-negative design overcomes healthcare-seeking selection bias by enrolling patients who present to the same clinic with identical acute clinical syndromes and partitioning them into lab-confirmed cases and test-negative controls.
- Information (Recall and Interviewer) Bias: Cases diagnosed with a serious malignancy or congenital anomaly frequently ruminate over past environmental or pharmacological exposures more thoroughly than healthy controls, inflating self-reported exposure prevalence among cases (differential recall bias). Nested case–control studies embedded within biobank cohorts eliminate recall bias by assaying archived pre-diagnostic blood samples or administrative prescription records collected years before symptom onset.
Frequently Asked Questions
Why is it necessary to use conditional logistic regression in an individually matched case–control study?
When controls are individually matched to cases on confounding factors such as age, sex, or clinic site (1:1 or 1:m matching), matching introduces selection bias toward the null if ignored, while fitting a separate intercept parameter αi for every matched pair in ordinary unconditional logistic regression causes the maximum-likelihood estimator of the odds ratio to converge to OR2 (Breslow & Day, 1980). Conditional logistic regression conditions on the sufficient statistic for each matched stratum, eliminating the nuisance intercepts and yielding an unbiased estimate of the adjusted odds ratio.
Can a case–control study directly measure disease incidence or prevalence?
A standard case–control study cannot calculate population disease incidence or prevalence on its own because the ratio of cases to controls is chosen arbitrarily by the investigator. However, if the known sampling fractions of cases (f1) and controls (f0) from a census-enumerated source population are available, population-level incidence rates and attributable fractions can be reconstructed mathematically via inverse-probability-of-sampling weights.
References
- Doll, R., & Hill, A. B. (1950). "Smoking and carcinoma of the lung: preliminary report." British Medical Journal, 2(4682), 739–748. https://doi.org/10.1136/bmj.2.4682.739
- Cornfield, J. (1951). "A method of estimating comparative rates from clinical data: applications to cancer of the lung, breast, and cervix." Journal of the National Cancer Institute, 11(6), 1269–1275. https://doi.org/10.1093/jnci/11.6.1269
- Mantel, N., & Haenszel, W. (1959). "Statistical aspects of the analysis of data from retrospective studies of disease." Journal of the National Cancer Institute, 22(4), 719–748. https://doi.org/10.1093/jnci/22.4.719
- Miettinen, O. (1976). "Estimability and estimation in case-referent studies." American Journal of Epidemiology, 103(2), 226–235. https://doi.org/10.1093/oxfordjournals.aje.a112220
- Breslow, N. E., & Day, N. E. (1980). Statistical Methods in Cancer Research, Volume I: The Analysis of Case-Control Studies. International Agency for Research on Cancer (IARC Scientific Publications No. 32). https://publications.iarc.fr/Book-And-Report-Series/Iarc-Scientific-Publications/Statistical-Methods-In-Cancer-Research-Volume-I-The-Analysis-Of-Case-Control-Studies-1980