How case-control studies work backwards from disease
A case control study starts with people who have a disease and comparable people who do not, then examines their earlier exposures. The comparison identifies associations, but its reliability depends on how the control group was drawn and how exposure histories were obtained (Jackson and Griffin, 2024; Andrade, 2022).
The design is used in medical research and outbreak investigation when following an entire population would be difficult. Its evidence is read through control selection, matching, measures of association and errors in recalled history. The examples come from original epidemiological investigations, methods papers and official field guidance rather than hypothetical risk estimates.
What a case control study is and how it works
A case control study defines participation by whether a person has already experienced the outcome under investigation. A case has the outcome. A control does not. The investigator compares the groups' exposure histories instead of waiting for new disease to develop (Andrade, 2022).
The inquiry runs backwards. Jackson and Griffin (2024) describe the same logic in the CDC field manual: people with disease and a comparable group without disease are identified before their exposure experiences are compared. Existing records or interviews may make that historical comparison possible (Andrade, 2022).
Cases, controls and exposure histories
A disease is one possible outcome. Andrade (2022) includes a clinical diagnosis, treatment resistance or an adverse effect among characteristics that can define a case. The comparison group has to be defined against that precise outcome, because changing the meaning of a case changes who qualifies as a control.
Exposures precede the outcome being examined. The CDC uses the term for behaviours, environmental conditions and other factors that might influence disease occurrence (Jackson and Griffin, 2024). Exposure can be reconstructed from interviews or existing medical records, with opportunities for error when historical details are incomplete (Andrade, 2022).
This structure does not establish the proportion of all exposed people who will eventually become ill. Participants were recruited because of their outcomes, so the proportion of cases in the assembled sample was set by the research design (Dicker, 2025). The sampling fractions do not describe disease incidence (Dicker, 2025).
How researchers select cases and controls
Controls represent the source population. In their 2019 account of observational study designs, Gail and colleagues describe population-based studies in which incident cases and non-cases must represent their respective source groups. Recruiting controls solely because they are convenient does not meet that condition.
Case ascertainment changes recruitment. A study based on hospital diagnoses can enrol a different set of patients from a register covering all new diagnoses in an area. In the earlier British lung-cancer investigation described by Peto and colleagues (2000), the cases were hospital patients and controls were other hospital patients matched on age. Those controls came from hospitals (Peto et al., 2000).
The source population
These are the people eligible to become cases. CDC guidance states that suitable controls reflect that population and provide an estimate of exposure within it (Jackson and Griffin, 2024). Selection based on the exposure being tested defeats that purpose: unexposed people chosen as controls will mechanically make exposure appear associated with illness (Jackson and Griffin, 2024).
Population controls and hospital controls may therefore answer subtly different questions. In the 1950 British smoking study, some control diagnoses were subsequently recognised as smoking-related (Peto et al., 2000). Peto and colleagues (2000) explain that those diagnoses affected the representation of smokers among controls and biased estimated relative risks downward. The distortion arises from which patients entered the controls (Peto et al., 2000).
Matching and overmatching
Matching selects comparable characteristics. It also changes who becomes a control, even when the matched characteristics are related to disease or exposure and were chosen to manage pre-existing differences between the groups. Mansournia, Jewell and Greenland (2018) explain that matching itself can create selection bias, including when the variable used is not a confounder.
The same authors distinguish individual matching from frequency matching, in which the distribution of a characteristic is made similar across the groups. Matching cases and controls on too many characteristics can leave fewer informative exposure contrasts. Their review warns that it may reduce precision and still require statistical adjustment for residual confounding (Mansournia et al., 2018).
Types of case control study designs
The sampling rules vary. Gail and colleagues (2019) distinguish studies recruiting from a general population or hospitals from designs drawing participants from an existing cohort. Although the studies may all compare exposure histories, the relation of the sampled controls to the underlying population determines the parameter estimated.
Matched and unmatched designs
An unmatched study has no such pairing. In a matched design, each case may be paired with a control, or the overall distributions of specified characteristics can be aligned by frequency matching (Mansournia et al., 2018). Matching changes the sample structure and must be incorporated into analysis.
The analysis must account for matching. The CDC states that case-control analysis must account for matching when it was part of the design (Dicker, 2025). Mansournia and colleagues (2018) also show why matching does not automatically balance confounders across exposure groups; it acts across groups defined by outcome.
Nested and case-cohort designs
A nested case-control study begins with a defined cohort, then samples controls from people still at risk when an individual becomes a case. Gail and colleagues (2019) describe controls time-matched to each event, with exposure information collected from the selected participants. The original cohort supplies the risk set and makes the sampling basis explicit.
Richardson (2004) evaluated incidence-density sampling for such designs using occupational cohort data. His analysis found biased relative-risk estimates under one previously proposed sampling programme, whereas the alternative evaluated in the paper produced unbiased estimates in that setting. The choice of sampling method therefore affects the estimate even if the exposure records themselves are accurate.
Case-cohort sampling differs. It includes cases from a cohort and samples a subcohort from the cohort as a whole (Gail et al., 2019). Labrecque and colleagues (2021) explain that these different sampling procedures determine whether a study can estimate an odds ratio, risk ratio or incidence rate ratio.
Advantages of case control studies and when they are useful
Case-control sampling begins with outcomes. In settings where diseases are rare and only a small fraction of people are affected, selecting existing cases can avoid following a very large number of unaffected people until enough cases appear. Mann (2003) identifies rare diseases as a setting in which this retrospective comparison is useful, and Gail and colleagues (2019) describe lower costs and shorter investigation times than collecting a new cohort from scratch.
The benefits depend on how much reliable information exists about the selected people. A fast interview-based comparison may be unable to reconstruct the relevant history with consistent accuracy; an existing cohort with recorded exposures may offer a different balance of efficiency and information quality (Andrade, 2022; Gail et al., 2019).
Rare diseases and outbreaks
Rare diseases require lengthy cohort follow-up. A conventional cohort may need to observe a very large number of people before sufficient outcomes arise, whereas a case-control investigation can recruit the already identified cases. Lewallen and Courtright (1998) describe research on risk factors for the rare eye cancer uveal melanoma as an example.
Outbreak investigations use case-control comparisons. Nielsen and colleagues (2023) investigated an Irish Salmonella Typhimurium cluster using cases from the outbreak and controls with other gastrointestinal illnesses. The investigators identified 18 outbreak cases and used a matched case-case design to examine chocolate product consumption. Its controls had other gastrointestinal illnesses (Nielsen et al., 2023).
Speed and sample efficiency
The CDC describes common control ratios. In field case-control investigations, one, two or three controls may be recruited for each case, with additional controls able to improve power when few people have become ill (Jackson and Griffin, 2024). Gains become small beyond three or four controls. The number is a design consideration, not a universal prescription for every dataset (Jackson and Griffin, 2024).
An existing source population can shorten recruitment and exposure measurement. Gail and colleagues (2019) describe population-based case-control work as less costly and time-consuming than launching a new cohort involving primary data collection. Sampling fewer records does not remove the need to specify which records qualified and when an exposure could have occurred.
What case control studies reveal about exposures and disease
The estimate describes an association. In a conventional case-control comparison, the odds that cases were exposed are compared with the exposure odds of controls selected under a specified sampling design. Lewallen and Courtright (1998) describe the odds ratio as this comparison and stress the need for a confidence interval when it is interpreted.
The sampling scheme changes the estimand. Labrecque and colleagues (2021) caution that designs using incidence-density sampling can have incidence-rate ratios as their estimands, even when investigators report an odds ratio from the sampled data. The statistical label cannot be separated from the route by which participants entered the study.
Odds ratios and their interpretation
The CDC describes an odds ratio as the odds of exposure in cases divided by the odds of exposure in controls (Dicker, 2025). A value above one means exposure was more common on the odds scale among cases; the interpretation concerns association within the comparison. Its estimate is ordinarily calculated from the cross-products of an exposure-by-outcome table (Dicker, 2025).
Confidence intervals describe precision. Lewallen and Courtright (1998) explain that the interval should accompany the odds ratio, because a point estimate alone gives a weak account of precision. In a sampled case-control study, population incidence and absolute risk cannot ordinarily be reconstructed directly from the chosen numbers of cases and controls (Dicker, 2025).
Labrecque and colleagues (2021) identify a further complication: case-cohort and incidence-density designs are not estimating the same population measure as every cumulative-sampling case-control design. Their paper examines cases in which the published measure and the valid estimand diverge. The sampling frame, follow-up period and eligibility for control selection belong beside any numerical result.
Smoking and lung cancer
Historical control groups reveal selection problems. Peto and colleagues (2000) compared work centred on 1950 with another study centred on 1990. Their report gives 1465 case-control pairs in the earlier investigation and 982 cases with 3185 controls in the later one. The periods involved different smoking histories and recruitment arrangements.
The 1950 controls were other hospital patients, chosen to match cases on age (Peto et al., 2000). Some had conditions later known to be associated with smoking (Peto et al., 2000). In those controls, smoking-related diagnoses meant that relative risks from the earlier comparison were too low (Peto et al., 2000).
The later comparison involved population and hospital controls whose smoking distributions were reported as closely similar. Peto and colleagues (2000) discussed changing smoking patterns when relating the two investigations to national death statistics. A separate account of coronary heart disease across populations addresses a different observational comparison, in which the composition of populations also affects what the results mean.
Published study comparisons
Large estimates can be imprecise. In Ireland, Nielsen and colleagues (2023) reported a matched odds ratio of 10.50 for consumption of at least one implicated chocolate product when comparing Salmonella outbreak cases with other gastrointestinal cases. The 95% confidence interval was 1.24–88.60, with p = 0.031. The wide interval limits the precision of that outbreak-specific estimate.
For a particular product, the association was not statistically significant: the study reported a matched odds ratio of 7.77 with a 95% confidence interval of 0.89–67.20 (Nielsen et al., 2023). That interval included one, so the particular comparison did not exclude no association at the conventional threshold. The two results concern distinct exposure definitions rather than independent replications.
Study | Participants or material | Comparison | Result and interpretive boundary |
Peto et al., 2000, earlier study | 1465 case-control pairs | Lung cancer cases and matched hospital controls | Hospital controls included diagnoses later recognised as smoking-related |
Peto et al., 2000, later study | 982 cases and 3185 controls | Lung cancer cases and hospital or population controls | Sampling populations and smoking histories differed from the earlier comparison |
Nielsen et al., 2023 | 9 analysed outbreak cases and 24 matched controls | Salmonella outbreak cases versus other gastrointestinal cases | Composite exposure mOR 10.50, 95% CI 1.24–88.60 |
Mésidor et al., 2026 | 418 studies screened, 23 included | Review of causal-inference methods in case-control research | Selected methodological literature, not a pooled disease effect |
Mésidor and colleagues (2026) searched for case-control papers applying specified causal-inference methods across the period March 2014 to March 2024. They identified 418 papers and included 23 in their methodological review. The review focused on papers applying causal-inference methods (Mésidor et al., 2026).
Disadvantages and biases of case control studies
Selection can distort the exposure contrast. Jackson and Griffin (2024) warn that selecting controls because they lack an exposure can create an apparent disease association even where none exists. The problem concerns the sampling mechanism and is not resolved by a more detailed statistical model of the selected participants.
Confounders affect exposure and outcome. Andrade (2022) explains that logistic-regression adjustment can address measured confounders but does not make an observational association automatically causal. Unmeasured confounding remains a plausible explanation when the relevant factors were never collected.
Selection bias and recall bias
Selection bias acts before analysis. The early lung-cancer hospital controls provide a documented case: the inclusion of patients with smoking-related illnesses altered the comparison with lung-cancer cases (Peto et al., 2000). The number of controls would not identify that error if their diagnoses and eligibility were left unexamined.
Recall may differ by disease status. People interviewed after an illness may reconstruct former behaviour differently from controls, while documents written before illness can offer another way to establish exposure (Andrade, 2022). The limitations of remembering past participation are also relevant to research on self-reported recreation participation, although that separate study addresses its own survey population.
Exposure definitions affect measurement. CDC field guidance describes how a person's recollection of food or other past events may be incomplete and how such errors can affect outbreak inquiries (Jackson and Griffin, 2024). Differences in memory between cases and controls can distort the comparison without changing the actual exposure histories.
Confounding and matching problems
Matching cannot guarantee comparable exposure groups. Mansournia and colleagues (2018) explain that matching operates across outcome groups and may itself induce selection bias. Their analysis also requires that statistical estimation reflect the particular matching procedure and residual confounding (Mansournia et al., 2018).
Andrade (2022) describes logistic regression as a way of estimating an adjusted exposure association in case-control data. Adjusted odds ratios still depend on what confounders were measured and whether the selected groups were representative. An association between education and self-rated health across countries illustrates a separate setting in which group characteristics complicate interpretation; it is not a case-control estimate.
Differences from cohort studies
Cohort studies classify participants by exposure or cohort membership before comparing outcomes, even when investigators reconstruct the sequence using historical records (Gail et al., 2019). Case-control selection generally begins with cases and a sampled comparison group, so it does not by itself supply the denominators needed to calculate attack rates (Mann, 2003; Dicker, 2025). The studies can address related scientific questions but supply different kinds of measurements.
Nested designs connect the approaches. Richardson (2004) examined controls sampled from an existing occupational cohort at the time cases occurred and demonstrated how the details of the sampling programme affected estimates. For such a design, the availability of risk sets changes what the resulting comparison can estimate.
What is proven about case control studies and what is not
Case-control studies can identify associations between past exposures and outcomes when cases and controls are recruited from a suitable source population. Their measurement is conditional on how controls were sampled, how matching was handled and how reliable the exposure histories are (Jackson and Griffin, 2024; Mansournia et al., 2018). The design alone does not establish causation (Andrade, 2022).
The same disease and exposure records can yield a misleading comparison if controls do not represent the source population that produced the cases (Jackson and Griffin, 2024).
The design's value for rare diseases is supported by established methodology. Mann (2003) describes it as suitable when outcomes are uncommon, while Gail and colleagues (2019) describe practical efficiency compared with collecting a new cohort. The narrower sample, however, prevents direct estimation of population incidence from a conventional case-control table (Dicker, 2025).
Sampling changes the target statistic. Labrecque and colleagues (2021) distinguish designs intended to recover odds ratios from incidence-density and case-cohort sampling arrangements with different estimands. Matching also imposes obligations on the analysis that remain even when cases and controls are similar in measured characteristics (Mansournia et al., 2018).
What a particular observed association means depends on how cases were detected, how exposure was measured and whether confounding has been addressed. In a 2026 methodological review, Mésidor and colleagues found that causal-inference techniques had been applied in a selected minority of the studies screened. They identified continuing methodological limitations, including how sampling bias and time-varying confounding are handled.
The odds ratio depends on its population. The Irish outbreak estimate concerned a specific chocolate exposure and a comparison with other gastrointestinal cases, while the British smoking research examined hospital and population controls under different historical conditions (Nielsen et al., 2023; Peto et al., 2000). Each result is informative within the population and design that produced it.
Frequently asked questions
Is every case-control study retrospective?
No. Case-control studies select people according to outcome status, then compare exposures (Jackson and Griffin, 2024). A nested design can sample cases and controls from an existing cohort with information recorded earlier, so the exposure data need not depend on later recollection (Gail et al., 2019). The defining feature is outcome-based selection, not the age of the records.
Can a case-control study prove causation?
No single observational case-control comparison automatically establishes cause and effect. An association can be influenced by confounders or by the way cases and controls were selected (Andrade, 2022; Mansournia et al., 2018). More elaborate causal-inference methods are being studied, but their validity depends on assumptions and on addressing sampling bias (Mésidor et al., 2026).
Can a case-control study measure disease incidence?
A conventional case-control sample does not directly measure disease incidence because numbers of cases and controls are chosen by the study design. The CDC explains that attack rates and a risk ratio cannot be computed from such a sample (Dicker, 2025). Cohort-based sampling schemes can support different rate-related estimands under specified conditions (Labrecque et al., 2021).
Does matching remove confounding?
No. Matching may make cases and controls similar on selected characteristics, but the groups are chosen by outcome. Mansournia and colleagues (2018) explain that matching can introduce selection bias and does not guarantee balance across exposure groups. Analysis must reflect the matching scheme and address remaining confounding (Dicker, 2025).
Why can an odds ratio differ from a risk ratio?
An odds ratio compares odds, while a risk ratio compares probabilities of an outcome. The CDC calculates an exposure odds ratio from sampled cases and controls, for which population attack rates generally cannot be recovered (Dicker, 2025). Labrecque and colleagues (2021) show that some control-sampling designs target rate or risk ratios instead of a conventional odds ratio.
Sources
- Mann CJ. Observational research methods. Research design II: cohort, cross sectional, and case-control studies. Emergency Medicine Journal. 2003;20(1):54–60. DOI 10.1136/emj.20.1.54.
- Jackson BR, Griffin PM. Designing and Conducting Analytic Studies in the Field. CDC Field Epidemiology Manual. 8 August 2024. https://www.cdc.gov/field-epi-manual/php/chapters/design-conduct-analyze-field-studies.html
- Gail MH, Altman DG, Cadarette SM, et al. Design choices for observational studies of the effect of exposure on disease incidence. BMJ Open. 2019;9(12):e031031. DOI 10.1136/bmjopen-2019-031031.
- Mansournia MA, Jewell NP, Greenland S. Case-control matching: effects, misconceptions, and recommendations. European Journal of Epidemiology. 2018;33(1):5–14. DOI 10.1007/s10654-017-0325-0.
- Labrecque JA, Hunink MGM, Ikram MA, Ikram MK. Do Case-Control Studies Always Estimate Odds Ratios? American Journal of Epidemiology. 2021;190(2):318–321. DOI 10.1093/aje/kwaa167.
- Mésidor M, Xu M, Diop A, et al. Use of causal inference methods in case-control studies: a methodology review. American Journal of Epidemiology. 2026;195(5):1438–1446. DOI 10.1093/aje/kwaf182.
- Andrade C. Research Design: Case-Control Studies. Indian Journal of Psychological Medicine. 2022;44(3):307–309. DOI 10.1177/02537176221090104.
- Peto R, Darby S, Deo H, et al. Smoking, smoking cessation, and lung cancer in the UK since 1950: combination of national statistics with two case-control studies. BMJ. 2000;321(7257):323–329. DOI 10.1136/bmj.321.7257.323.
- Nielsen CS, Garvey P, Cormican M, et al. Investigation of a monophasic Salmonella Typhimurium outbreak linked to chocolate products as part of wider international outbreak: A matched case-control study, Ireland, 2022. Public Health Challenges. 2023;2(4):e116. DOI 10.1002/puh2.116.
- Dicker RC. Analyzing and Interpreting Data. CDC Field Epidemiology Manual. 6 January 2025. https://www.cdc.gov/field-epi-manual/php/chapters/analyze-interpret-data.html
- Lewallen S, Courtright P. Epidemiology in practice: case-control studies. Community Eye Health. 1998;11(28):57–58.
- Richardson DB. An incidence density sampling program for nested case-control analyses. Occupational and Environmental Medicine. 2004;61(12):e59. DOI 10.1136/oem.2004.014472.
Related
How Accurate Are Self-Reported Recreation Participation Rates?
Self-Rated Health and Education: Findings from 69 Countries
Seven Countries Study: Coronary Heart Disease Across Populations
Prophylactic Inoculations and Poliomyelitis: 620 Victorian Cases
What internal and external validity say about trusting a result
The placebo effect and how expectations can change how we feel
Chinese tourism research: questionnaire barriers in Kanas