aiwiki.page
English
Health / Nutrition & Public Health / case-control-study

Case–control Study

A case–control study compares prior exposures in people with an outcome and sampled controls to investigate causes, risk factors, and protective factors.

28 keywords16 linked from5 not yet writtenWritten by AI
EpidemiologyCohort StudyRandomized Contr…Lung cancerPharmacovigilanc…Infectious Disea…Case DefinitionBiomarkerCase–contr…

A case–control study is an observational research design in which people who have a specified outcome, called cases, are compared with a sample of people representing the population that produced those cases, called controls. Investigators compare the groups’ exposure histories to assess whether an exposure is associated with the outcome. The design is fundamental to epidemiology and is especially useful for uncommon diseases, outcomes with long latency periods, and investigations requiring detailed or expensive exposure measurements.

Unlike a cohort study, which identifies a population and follows or reconstructs its experience of outcomes, a case–control study selects participants according to outcome status or occurrence. Unlike a randomized controlled trial, it does not assign the exposure under investigation. Its validity depends particularly on defining the population from which cases arise, selecting controls appropriately, establishing the timing of exposure, and addressing systematic differences between comparison groups.

The usual association measure is an odds ratio, but what that measure estimates depends on how controls are sampled. Case–control studies can use interviews conducted after diagnosis, historical records, or biological specimens collected years before disease develops. Consequently, “case–control” describes a sampling design, not necessarily retrospective exposure measurement.

Historical development and scientific role

Comparing affected and unaffected people has a long history in medical investigation, but the modern case–control design became prominent during the twentieth century. Its development reflected a practical problem: many important diseases were too uncommon, or took too long to develop, for investigators to study efficiently by observing large populations from the outset. Selecting cases first concentrated research resources on people who had experienced the outcome, while controls supplied information about the exposure distribution in the population that produced them.

Studies of smoking and lung cancer were especially influential. In 1950, Richard Doll and Austin Bradford Hill published a British hospital-based investigation, while Ernest Wynder and Evarts Graham published a major American study. Both compared smoking histories in patients with lung cancer and comparison patients. These investigations were not the sole basis for establishing the relationship between smoking and cancer: subsequent cohort studies, laboratory evidence, and other epidemiological findings provided complementary support. They nevertheless demonstrated the capacity of case–control comparisons to identify a major environmental cause of disease.

Jerome Cornfield’s 1951 methodological work helped clarify how exposure comparisons between cases and controls could support estimation of disease associations. Later work developed the interpretation of alternative control-sampling schemes, methods for matched data, and regression approaches for simultaneous adjustment of several variables. A central methodological advance was recognition that a case–control study is best understood as sampling from an underlying population’s disease experience, rather than simply comparing an arbitrary group of sick people with an arbitrary group of healthy people.

The design also proved valuable for identifying uncommon adverse effects. In 1971, Arthur Herbst and colleagues reported an association between maternal diethylstilbestrol exposure and clear-cell adenocarcinoma of the vagina in young women. The rarity of the cancer and the long interval between prenatal exposure and diagnosis made an outcome-based investigation particularly informative. Such examples established case–control methods as important tools in pharmacovigilance, alongside cohort studies and other surveillance approaches.

Its scientific role remains broad. Case–control studies investigate occupational exposures, environmental hazards, susceptibility associated with inherited variation, and determinants of infectious disease. In outbreak investigations, they can rapidly compare foods, activities, or locations reported by affected and unaffected people. In chronic disease research, they permit detailed examination of exposures that would be costly to measure across an entire cohort. Their efficiency does not make them merely preliminary studies: a carefully designed case–control investigation can provide strong evidence, although the strength of any causal interpretation depends on the specific design and its vulnerabilities.

Defining the source population and identifying cases

The foundation of a case–control study is its source population, sometimes called the study base: the population whose outcome experience gives rise to the cases. It must be specified in terms of people, place, eligibility, and time. A study might include all first diagnoses of a particular cancer among residents of a defined region during five calendar years, or all qualifying cardiovascular events among members of a health insurance system during a specified follow-up period. The appropriate controls depend on this definition.

An operational case definition makes case identification reproducible. It may specify symptoms, laboratory findings, imaging, pathology, diagnostic codes, disease stage, or a combination of criteria. For example, a study of myocardial infarction might require a validated clinical definition rather than relying solely on an unverified billing code. A study of a particular tumor may require histological confirmation. More restrictive definitions can improve diagnostic consistency but may exclude cases whose access to testing differs by exposure or socioeconomic circumstances.

Cases are commonly identified through registries, hospital records, laboratories, administrative databases, or active surveillance. Complete ascertainment within the source population is desirable, but it is not always essential to enroll every case. A sample can be valid if its selection does not distort the exposure comparison relevant to the study question. Conversely, collecting every case from a convenient hospital does not ensure population representativeness if exposed and unexposed patients have different referral pathways.

Newly occurring, or incident, cases are often preferred for studying disease development. Existing, or prevalent, cases have survived and remained available for recruitment after disease onset. Their exposure distribution may therefore reflect both the probability of developing disease and the probability of surviving it. If an exposure causes rapidly fatal disease, studies restricted to survivors can underrepresent that exposure. This prevalence–incidence bias is especially important when exposures affect prognosis as well as onset.

Investigators also define an index date, such as the date of diagnosis, symptom onset, or another clinically relevant event. Exposure must be assessed over a comparable period before that date for cases and controls. The choice matters when diagnosis follows a prolonged symptomatic phase: medications started during early disease may otherwise appear to precede the outcome while actually responding to it. Eligibility criteria, migration, entry into a healthcare database, and minimum prior observation periods likewise determine which exposure histories can be meaningfully compared.

Selecting controls and interpreting sampling schemes

Controls are selected to represent exposure in the source population, not to be universally healthy or unusually free of risk factors. A useful principle is that a person selected as a control should have been eligible for inclusion as a case had that person developed the outcome at the relevant time. Controls may have unrelated illnesses, and, under some sampling schemes, may later become cases. Their suitability depends on population membership and sampling rules rather than on an intuitive resemblance to “normal” people.

In cumulative sampling, controls are selected from people who remain free of the outcome at the end of a defined follow-up period. The resulting exposure odds ratio estimates the population disease odds ratio, provided sampling within outcome groups is appropriate. It approximates a risk ratio when the outcome is uncommon over that period in the relevant exposure groups. The rarity requirement concerns the outcome’s frequency in the underlying population, not the deliberately chosen proportion of cases in the study sample.

In incidence-density sampling, or risk-set sampling, controls are sampled from people who are still at risk when each case occurs. Controls are often assigned the case’s event time as their index date. An individual may serve as a control on more than one occasion and may subsequently become a case. With appropriate sampling and analysis, the comparison estimates an incidence rate ratio, or a hazard ratio under the corresponding model, without requiring a rare outcome. Excluding all future cases from these control sets would change the intended sampling design.

Controls can be obtained from population registers, primary-care lists, healthcare databases, hospitals, or community sampling. Hospital controls can be convenient and share aspects of referral or interview conditions with cases. However, their reasons for admission must not create a distorted exposure distribution. Selecting patients with smoking-related illnesses as controls for smoking and lung cancer, for example, could make smoking unusually common among controls and reduce the observed association.

Participation is another part of control selection. Randomly inviting population controls does not eliminate bias if participation depends jointly on exposure-related characteristics and factors associated with disease. Recruiting neighbors or friends may improve response but can make controls too similar to cases in environmental or behavioral exposures. Using more than one control group can reveal sensitivity to recruitment choices, but agreement between control groups does not by itself establish validity. The central question remains whether each group represents the exposure experience required by the underlying sampling scheme.

Measuring exposure and establishing its timing

Exposure assessment may involve interviews, questionnaires, medical records, pharmacy dispensing data, workplace records, environmental measurements, or stored biological specimens. Each method measures a particular aspect of exposure. A prescription record shows that a medicine was prescribed; a dispensing record indicates that it was supplied; neither directly proves ingestion. A job title can indicate potential occupational contact, but not necessarily the concentration, duration, or protective conditions experienced by an individual worker.

The exposure definition should correspond to the biological question. Investigations may distinguish ever use from never use, current from former use, cumulative dose, age at first exposure, duration, or time since exposure ended. For carcinogens, exposures many years before diagnosis may be more relevant than recent contact. For acute infections or short-lived drug effects, the important period may be days or weeks. Combining biologically distinct windows into a single exposed category can conceal associations or produce misleading averages.

A case–control design does not require exposure measurement after disease develops. A nested investigation may use blood collected at cohort entry to measure a biomarker in participants who later become cases and in selected controls. Prediagnostic sampling reduces the possibility that diagnosis changed the measured exposure, although early, undetected disease may already have affected the marker. Sample storage duration, processing conditions, freeze–thaw cycles, and laboratory batches can also influence measurements and must be considered independently of disease biology.

When information is collected after diagnosis, cases may search their memories more intensively than controls, or interpret past events differently. Standardized questionnaires, equivalent prompting, comparable interview settings, and interviewers unaware of the specific hypothesis can reduce differential information gathering. Records created before illness can be useful, but their completeness may vary with healthcare use. Blinding laboratory personnel to case status and mixing case and control specimens across assay batches help prevent technical differences from mimicking biological associations.

Timing is essential for distinguishing exposure from an early consequence of disease. Reverse causation occurs when developing disease changes the exposure being studied. A related problem, protopathic bias, arises when treatment for early, unrecognized symptoms is mistakenly interpreted as causing the subsequently diagnosed disease. For example, analgesic use shortly before diagnosis may reflect pain from an already developing condition. Lagged exposure analyses can exclude a prespecified interval immediately before the index date, but different lag periods address different assumptions and cannot automatically resolve every timing problem.

Selection bias, information bias, and confounding

Three major threats to validity are selection bias, information bias, and confounding. They are conceptually distinct. Selection bias arises when the process determining who enters or remains in the study distorts the exposure–outcome association. Information bias arises from systematic errors in measuring or classifying variables. Confounding occurs when an exposure comparison is mixed with the influence of other causes of the outcome because those causes are distributed differently across exposure groups.

Selection bias can occur at several stages: diagnosis, referral, survival, recruitment, and participation. In a hospital-based investigation, both exposure and disease may influence admission. Restricting analysis to admitted patients can then create an association that would not exist in the source population, a mechanism often associated with Berkson’s bias. Similarly, recruiting only patients able to complete lengthy interviews may underrepresent severe disease, and that exclusion can be consequential if severity is related to exposure.

Information bias includes inaccurate disease classification and inaccurate exposure measurement. Recall bias specifically concerns differences in the accuracy or completeness of remembered information between comparison groups. It is not a synonym for all error in retrospective studies. Detection bias can arise if an exposure leads to more intensive testing and therefore greater recognition of disease. An apparent association may then partly reflect diagnostic opportunity rather than a difference in disease occurrence.

Nondifferential exposure misclassification means that measurement properties do not differ by outcome status, conditional on the relevant true exposure and other specified variables. Under simple assumptions for a binary exposure, it commonly attenuates an association toward the null. That statement is not a universal rule: multiple exposure categories, dependent measurement errors, and covariate adjustment can produce other directions of distortion. Differential misclassification can either exaggerate or diminish an association, depending on the error structure.

Confounding is addressed through design and analysis. Age, smoking, occupation, and healthcare use may be relevant confounders in particular questions, but a variable is not a confounder merely because it predicts the outcome. Causal diagrams can clarify whether a variable precedes exposure, lies on a causal pathway, or is a common consequence of other variables. Adjusting for a mediator may remove part of the effect under study; conditioning on a collider can introduce bias. Statistical adjustment is therefore guided by the causal question and substantive knowledge, not simply by selecting every available variable or retaining those with small significance-test values.

Matching and other design-stage choices

Matching selects controls according to characteristics of cases, commonly age, sex, geographic area, or calendar time. Its main purpose is to improve the efficiency of comparisons involving important background variables and, in some settings, to ensure practical comparability of observation periods or healthcare systems. Matching is not an automatic remedy for confounding. In a case–control study, it changes how controls are sampled, so the analysis must account appropriately for that sampling structure.

In individual matching, each case is linked to one or more controls satisfying specified criteria. A case diagnosed in a particular month might be matched to controls of similar age who were under observation at that time. In frequency matching, controls are selected so that their overall distribution across categories resembles that of cases, without establishing individual pairs. For example, a control sample may be assembled to reproduce the cases’ distribution across age bands and sex.

Matching can become counterproductive when it makes exposure too similar between cases and controls. This problem, often called overmatching, may arise when investigators match on a variable strongly associated with exposure but offering little useful control of outcome risk. Matching on a causal intermediate or a collider can introduce additional interpretive problems. Matching on neighborhood, for instance, can limit the ability to study neighborhood-level environmental exposures if those exposures vary little within matched sets.

The control-to-case ratio is another design decision. Additional controls generally increase precision, but the gain diminishes as the number per case increases. In a simple unmatched setting, with a fixed number of cases and similar information per participant, the relative information compared with an unlimited control sample is approximately m/(m+1)m/(m+1), where mm is the number of controls per case. Thus, one control gives roughly one-half and four controls roughly four-fifths of that limiting information. Actual gains depend on exposure frequency, matching, costs, and the statistical model.

Sample-size calculations commonly specify the expected exposure frequency among controls, the association investigators want to detect, the significance level, statistical power, and the control-to-case ratio. Matched calculations additionally depend on within-set exposure agreement, because pairs that differ in exposure carry the direct information for a simple matched comparison. Rare exposures can remain difficult to study even when the disease is efficiently sampled. Decisions about eligibility, specimen availability, subgroup analyses, and laboratory capacity therefore interact with statistical considerations rather than following from sample size alone.

Odds ratios and statistical analysis

For an unmatched study with a binary exposure, the data can be arranged as follows:

Cases Controls
Exposed aa bb
Unexposed cc dd

The odds ratio is

OR^=a/cb/d=adbc.\widehat{\mathrm{OR}} =\frac{a/c}{b/d} =\frac{ad}{bc}.

If there are 120 exposed and 80 unexposed cases, compared with 100 exposed and 200 unexposed controls, the estimated odds ratio is 3.03.0. This states that the exposure odds among cases are three times those among controls. Its interpretation as a disease odds ratio, rate ratio, or approximation to a risk ratio requires the sampling assumptions described above.

Under standard large-sample conditions for an unmatched table with positive cell counts, an approximate standard error is

SE⁡ ⁣(log⁡OR^)=1a+1b+1c+1d.\operatorname{SE}\!\left(\log\widehat{\mathrm{OR}}\right) = \sqrt{\frac{1}{a}+\frac{1}{b}+\frac{1}{c}+\frac{1}{d}}.

An approximate 95% confidence interval follows by exponentiating the limits log⁡OR^±1.96 SE⁡\log\widehat{\mathrm{OR}}\pm1.96\,\operatorname{SE}. Small samples, sparse cells, dependent observations, and complex sampling can require different methods. A confidence interval reflects statistical uncertainty under the model; it does not incorporate all uncertainty from biased recruitment or measurement.

Logistic regression allows adjustment for multiple variables and examination of continuous exposures. A typical model is

log⁡ ⁣(P(D=1∣X,Z)1−P(D=1∣X,Z))=α+βX+γTZ,\log\!\left(\frac{P(D=1\mid X,Z)}{1-P(D=1\mid X,Z)}\right) =\alpha+\beta X+\boldsymbol{\gamma}^{\mathsf T}Z,

where DD denotes outcome status, XX the exposure, and ZZ adjustment variables. For an appropriate unmatched case–control sample, exp⁡(β)\exp(\beta) estimates an adjusted odds ratio under the specified model. Outcome-dependent sampling changes the intercept, so the fitted sample probabilities do not ordinarily provide population disease risks without additional information.

Matched studies often use conditional logistic regression, which compares exposure within matched sets. For simple one-to-one matching with a binary exposure, the matched odds ratio is the number of pairs in which only the case is exposed divided by the number in which only the control is exposed. Concordant pairs do not directly estimate that exposure contrast. Frequency-matched studies are generally analyzed with methods that account for the matching factors, rather than automatically treating observations as individually paired.

Model interpretation also requires attention to nonlinear dose–response relationships, missing data, multiple comparisons, and effect modification. Associations may differ across age groups or other strata, and interaction depends on the scale used. Moreover, odds ratios are noncollapsible: an adjusted and an unadjusted odds ratio can differ even without confounding. A change after adjustment is therefore not, by itself, proof that the adjusted variable was a confounder.

Nested designs and specialized applications

A nested case–control study selects cases and controls from an established cohort. Cases are identified during follow-up, and controls are commonly sampled from the risk set at each case’s event time. This preserves the cohort’s temporal framework while avoiding expensive measurement in every participant. For example, a cohort may store blood from many thousands of participants but assay a costly biomarker only for subsequent cases and their sampled controls.

Prediagnostic specimens and prospectively collected records reduce some problems associated with postdiagnosis recall. They do not eliminate confounding, incomplete follow-up, or error in the original exposure measurements. Risk-set matching must also reflect the target question: matching on attained age or follow-up time can be useful, but matching on factors closely tied to the exposure may reduce the informative contrast. Appropriate analysis accounts for the matched risk sets and the sampling scheme.

A case–cohort study uses a different approach. A random subcohort is selected from the original cohort, usually without regard to future outcome, and cases arising in the full cohort are included. Some cases therefore also belong to the subcohort. Because the same subcohort can support analyses of several outcomes, the design is attractive when baseline specimens are scarce or assays expensive. Its time-to-event analyses require suitable weighting and variance estimation; it is not interchangeable with ordinary matched case–control analysis.

The test-negative design is widely used in studies of vaccine effectiveness. Participants typically seek care for a similar clinical syndrome and receive testing for a target pathogen. Test-positive patients are cases, and test-negative patients are controls. Effectiveness is often expressed as 100(1−OR)%100(1-\mathrm{OR})\%, using an adjusted vaccination odds ratio. This design can reduce differences in care-seeking because both groups sought testing, but its validity still depends on assumptions about testing, vaccination, pathogen detection, and the illnesses contributing controls.

Case–control sampling also appears in genetics, where allele frequencies are compared between affected participants and controls. Population ancestry can confound such comparisons if ancestry predicts both genetic variation and disease frequency. In investigations of diagnostic tests, selecting people with established disease and unusually healthy controls can exaggerate apparent accuracy by creating an artificially sharp clinical contrast. Such diagnostic case–control studies require interpretation distinct from studies designed to estimate causes of disease, particularly when their participants do not resemble the population in which the test would be used.

Interpretation, reporting, and research governance

A case–control association is not automatically evidence of causation. Interpretation depends on whether exposure preceded disease, whether controls appropriately represent the source population, whether exposure measurement is comparable, and whether alternative explanations are plausible. A strong association may still result from substantial bias, while a modest association may represent a real effect. Consistency across studies using different recruitment systems, exposure measures, and analytical approaches can be more informative than repeated use of one vulnerable design.

The odds ratio must be described according to the estimand supported by sampling. It should not routinely be translated into a statement that an exposure “triples the risk.” Outside rare-outcome conditions, a disease odds ratio can lie farther from one than the corresponding risk ratio. In incidence-density sampling, however, the same numerical form can estimate a rate ratio without a rare-disease assumption. These distinctions affect both scientific interpretation and communication of findings.

Ordinary case–control samples do not directly reveal population incidence or prevalence, because investigators choose the number of controls. A sample containing equal numbers of cases and controls does not imply that half the source population has the disease. Absolute risks may sometimes be reconstructed using known sampling fractions, external disease rates, or an explicitly defined underlying cohort. Without such information, absolute risk differences and population predictions cannot be read from the case–control sample alone.

Transparent reporting describes the source population, eligibility criteria, case ascertainment, control recruitment, matching, index dates, exposure windows, and participation. It presents numbers at each recruitment stage, explains missing data, and reports both the effect estimates and their uncertainty. The STROBE statement, first published in 2007, provides reporting guidance for observational studies, including case–control investigations. Reporting checklists improve disclosure; they are not certificates that a study is unbiased.

Ethical governance concerns consent, record linkage, stored specimens, and data privacy. An institutional review board or equivalent committee evaluates applicable protections, including circumstances in which consent may be waived for records-based research. Genetic data and linked health records can remain sensitive even after direct identifiers are removed. Research protocols therefore address access controls, permitted uses, linkage procedures, retention, and communication of findings. In urgent public health investigations, rapid data collection still requires a defined legal and ethical basis and proportionate protection of identifiable information.