Can one person be both a control and a case?

Yes, at different times, as B demonstrates.

Does nested mean the study must be small?

No. It describes how cases and controls are selected within a cohort, not a fixed sample size.

AI Glossary / Statistics and probability / Advanced

Nested case - control study

Definition

A nested case - control study selects cases and controls from a defined cohort. In the common risk-set design, each case is paired with controls sampled from cohort members who are still at risk when that case occurs. Detailed measurements can then be collected for selected members rather than for the entire cohort.

What it means

The cohort supplies the membership list and follow-up history; the nested sample supplies more intensive information. A control represents someone eligible to become a case at a particular time. It does not mean someone guaranteed to remain event-free forever. The design can reduce measurement costs while retaining the timing needed for an appropriate event-time comparison.

How it works

Define the cohort, event, follow-up period, matching rules, and control-sampling procedure. At each case time, identify the eligible risk set and sample controls. Preserve the matched-set identifiers and the information available by that time. Conditional logistic regression is commonly used; under the corresponding Cox proportional-hazards model, it estimates hazard-ratio parameters. Other targets and sampling variants require appropriate analyses.

Three practical examples

A cohort study stores baseline specimens but cannot afford to assay every specimen. Researchers assay cases and sampled controls, selecting controls from the relevant risk sets.
A workplace cohort maintains event records for all workers while reconstructing detailed historical exposure only for selected cases and controls. Exposure reconstruction must use comparable time windows.
An AI-assisted records project uses detailed document review for a nested sample rather than every cohort member. Reviewers distinguish information available before the event from information created during its investigation.

Think of it this way

Imagine a long-running register of participants. Whenever an event occurs, a snapshot identifies who was still eligible at that moment. Controls come from that snapshot, not from a list chosen later using knowledge of who ultimately avoided the event.

Worked example

Consider an original cohort of eight members, A through H, all entering follow-up at time zero. A becomes a case at time 2, D at time 4, and B at time 6. H leaves observation event-free at time 3. The other members remain observed and event-free through time 6. For simplicity, assume no other matching restrictions and sample two distinct controls per case.

At time 2, all eight members are at risk. Excluding case A leaves seven eligible controls: B, C, D, E, F, G, and H. There are 21 possible unordered pairs, each equally likely under simple random sampling. Suppose B and C are selected. Each eligible member has probability 2/7 of selection for this set.

At time 4, A has already experienced the event and H has left observation. The risk set is B, C, D, E, F, and G. Excluding case D leaves five eligible controls. There are ten possible pairs, and each eligible member has probability 2/5 of being selected. Suppose B and E are selected. B can be selected again because prior control status does not remove an otherwise eligible member under this standard design.

At time 6, the risk set is B, C, E, F, and G. B is now the case. The four eligible controls are C, E, F, and G, giving six possible pairs and selection probability 1/2 for each member. Suppose F and G are selected. B's earlier appearance as a control remains valid: B was event-free at those earlier case times.

The resulting sets contain nine appearances: three cases and six control appearances. They represent seven distinct people, A through G. Counting nine appearances as nine independently sampled people would misdescribe the design. H was eligible in the first risk set but happened not to be selected; its later departure does not retroactively remove that eligibility.

To illustrate the matched analysis, assign a fixed binary exposure measured at baseline: A, C, and F are exposed; B, D, E, and G are unexposed. The three observed sets therefore contain an exposed case with one exposed and one unexposed control; an unexposed case with two unexposed controls; and an unexposed case with one exposed and one unexposed control.

Under a simple conditional model with exposure hazard ratio theta, their likelihood contributions are theta/(2theta+1), 1/3, and 1/(theta+2). At theta = 1, the combined likelihood is 1/27. At theta = 2 it is 1/30, and at theta = 1/2 it is also 1/30. The middle set contributes no exposure contrast because all three members have the same exposure. These calculations illustrate conditioning on sets, not a reliable substantive estimate from three cases.

Illustrative case study

This fictional case concerns Willow Records, an imaginary research team using AI to organize cohort documents. Its first sampling script removes anyone previously chosen as a control. A reviewer notices that the intended design allows repeat eligibility and corrects the script before sampling.

The team records risk-set membership, selection probabilities, case times, and duplicate appearances. It also limits exposure extraction to the specified pre-event window for both cases and controls. The case illustrates study construction, without asserting a health effect.

Common misconceptions and mistakes

Selecting only people who remain event-free to the end changes the control-sampling design. A later case can be a valid earlier control.

Likewise, excluding previously selected controls is a modified design, not a harmless deduplication step. Published methods address such modifications, but their sampling probabilities and analysis must agree. Matching also does not automatically eliminate confounding or turn an observational study into a randomized experiment.

Business use

This design is useful when expensive review or measurement is the main constraint and cohort follow-up is available. Budget for maintaining the full membership and event register even when detailed measurements use a subset. If the aim is absolute risk or AI prediction calibration, do not treat the selected case proportion as the cohort event probability; sampling-aware methods and additional cohort information may be needed.

Check your understanding

Why is H ineligible at time 4?

Answer: Follow-up ended at time 3.

Why is B still an eligible control at time 4?

Answer: B has not yet experienced the event, even though B later becomes a case.

Related terms

Cohort study
Hazard ratio
risk-set sampling
incidence-density sampling
conditional logistic regression
matching
selection probability
case-cohort study

Learn next

Reconstruct the three eligible-control lists and compare the standard design with a rule excluding previous controls. Identify which selection probabilities would change.

Sources

López-Nieto-Veitch et al., On the estimation of inclusion probabilities for weighted analyses of nested case control studies, author preprint, introduction and section 2.1: https://arxiv.org/html/2601.04066v1 . Shin and Saegusa, Nested case-control sampling without replacement, published author abstract: https://pubmed.ncbi.nlm.nih.gov/39235702/ . Cohort examples and Willow Records are original illustrations.

Previous: NER model Next: Nested cross-validation