Must controls remain outcome-free forever?

No. In some valid designs they are sampled while at risk and may become cases later.

Can a case-control study establish causation by itself?

An association requires careful causal reasoning and assumptions; the design does not automatically eliminate confounding or bias.

Does selecting equal case and control numbers mean the outcome affects half the population?

No. That fraction reflects the study's sampling choices.

AI Glossary / Statistics and probability / Intermediate

Case-control study

Definition

A case-control study selects people or other units according to whether they have an outcome of interest, then compares their prior exposures or characteristics. Cases have the defined outcome; controls are selected to represent the relevant source population under the sampling design. The design is useful for studying uncommon outcomes without measuring exposure in an entire cohort.

What it means

The defining feature is selection by outcome, not simply looking backward in time. Exposure information may come from interviews, earlier records or samples collected before the outcome occurred. Controls need to provide an appropriate comparison for the population that generated the cases. They are not just convenient healthy volunteers. The way controls are sampled determines what the resulting association estimates, and poor selection can create bias that additional statistical adjustment cannot necessarily repair.

How it works

Define the outcome, eligibility criteria, source population and relevant exposure period. Ascertain cases consistently and choose a control-sampling strategy suited to the question. Measure exposures comparably in both groups, ideally without knowledge that could distort assessment. Analyze exposure differences using methods appropriate to the design, often odds ratios and logistic regression. Account for matching or other sampling features rather than ignoring them. Investigate confounding, selection bias and measurement error. Because the numbers of cases and controls are typically chosen by design, their proportions in the dataset do not directly reveal population outcome risk.

Three practical examples

Public-health investigation: compare earlier exposures among people with a defined illness and suitable controls to identify plausible contributors.
Product reliability: compare failed components with appropriately sampled nonfailed components from the relevant production population and operating period.
Security research: examine previous account characteristics among confirmed compromise cases and controls sampled from accounts at risk, while preventing post-event information from leaking into the exposure assessment.

Think of it this way

Think of starting with incidents and asking what was different beforehand. The comparison group must come from the same opportunity for an incident to occur. Comparing damaged items from one environment with undamaged items from another may mainly reveal the difference between environments.

Worked example

Suppose an illustrative unmatched study selects 100 cases and 100 controls. Sixty cases were exposed and 40 were unexposed; among controls, 30 were exposed and 70 were unexposed. The exposure odds among cases are 60/40 = 1.5, and among controls they are 30/70, approximately 0.4286. Their ratio is (60 × 70)/(40 × 30) = 3.5. This is the sample exposure odds ratio. The tempting calculation 60/(60 + 30) gives two-thirds, but it is not the population risk of the outcome among exposed people: the investigator deliberately selected the case and control counts. If a comparable second control sample doubled both control cells to 60 and 140, the odds ratio would remain (60 × 140)/(40 × 60) = 3.5, while the naive exposed-case fraction would change to one-half. That contrast shows why a study-selected denominator cannot be treated as a population risk denominator. The example omits confounding and sampling uncertainty solely to clarify the arithmetic.

Illustrative case study

Fictional case: Copperline Devices investigates a rare battery failure. Investigators define the failure consistently, identify the production population and sample controls with relevant opportunities for failure. They obtain exposure information from manufacturing records created before the failures and examine operating conditions as potential confounders. The team treats an elevated odds ratio as evidence requiring further investigation, not proof that a component caused the failures. This fictional case describes an observational design and makes no claim about an actual product defect.

Common misconceptions and mistakes

Case-control does not mean randomized controlled trial: exposure is usually observed, not assigned. An odds ratio is not generally a risk ratio. Under particular sampling schemes and assumptions it estimates an odds ratio, approximates a risk ratio when appropriate, or estimates a rate ratio with incidence-density sampling; the design matters. Controls selected from a risk set can later become cases. Matching does not automatically remove confounding and must be handled in analysis. Recall bias is a concern when memories differ between groups, but it is not inevitable when reliable pre-existing exposure records are available.

Business use

Businesses can adapt outcome-based sampling to investigate rare failures, complaints or security incidents efficiently. Clearly define who could have become a case and when. Avoid choosing controls solely because they are easy to access or differ in follow-up opportunity. When AI is used to classify cases or extract exposures, validate errors separately in the comparison groups and preserve event timing. If a predictive model is trained on deliberately enriched case-control data, its raw probabilities may require correction or calibration for deployment prevalence; classification performance alone does not solve that issue.

Check your understanding

Q: Why did doubling the control counts leave the worked example's odds ratio unchanged?

Answer: A: Both control exposure categories were multiplied by the same factor, leaving their exposure odds unchanged. The apparent outcome fraction changed because its denominator depended on the chosen sampling ratio.

Related terms

Odds ratio
Cohort study
case definition
control selection
source population
incidence-density sampling
matching
confounding and selection bias

Learn next

construct the same two-by-two table from a complete cohort and compare the available risk and odds measures. Then study how sampling controls at the time each case occurs changes the target of inference, and why matched analyses use information differently from unmatched analyses.

Sources

Sources: Harrell, statistical glossary: https://hbiostat.org/glossary/ ; CDC Field Epidemiology Manual, study design and source-population controls: https://www.cdc.gov/field-epi-manual/php/chapters/design-conduct-analyze-field-studies.html ; Pearce, “What does the odds ratio estimate in a case-control study?”: https://pubmed.ncbi.nlm.nih.gov/8144304/ ; CDC, interpretation of odds ratios: https://www.cdc.gov/field-epi-manual/php/chapters/analyze-interpret-data.html .

Previous: Case-based reasoning Next: CAT