03 · What You Need to Know
How to Design a Nested Case-Control Study Within a Registry
What Exactly Is a Nested Case-Control Study?
A nested case-control study is a case-control investigation conducted within a previously defined cohort. The source cohort provides the population from which cases arise and from which eligible controls are selected.
Rather than measuring detailed exposure information for every cohort member, researchers obtain or analyze that information for identified cases and a sampled set of controls.
The approach can be particularly efficient when outcomes are relatively uncommon or exposure assessment is costly. It is not restricted to rare outcomes, however, and its efficiency depends on the number of cases, controls, and measurements required.
The word nested is important. Cases and controls are not simply recruited from the same institution or database. They must arise from a defined underlying cohort with an interpretable relationship between follow-up time, case occurrence, and control eligibility.
Does Every Registry Contain a Suitable Source Cohort?
No. A registry may contain extensive records without providing the information necessary to reconstruct a valid cohort.
A suitable source cohort requires identifiable eligibility criteria, an appropriate entry point, and a period during which participants can be observed for the outcome.
Researchers should be able to establish when individuals entered observation, whether they were eligible and outcome-free at the relevant starting point, and when they experienced the event, were censored, or otherwise left observation.
A registry containing only people who already experienced the outcome is generally insufficient to construct a nested case-control study because it does not identify the population at risk from which controls should be drawn.
Likewise, a database containing cases and a convenient group of noncases without reliable follow-up information does not automatically become a nested case-control study.
Where the necessary longitudinal structure exists, the registry may also support a retrospective cohort analysis. The choice between analyzing the full cohort and sampling controls depends partly on measurement costs, available information, and the intended estimand.
How Are Cases Identified Within the Cohort?
Cases are cohort members who experience the prespecified incident outcome during eligible follow-up.
Researchers need a defensible outcome definition and reliable event dates. Depending on the study, outcomes might include disease onset, hospitalization, a particular complication, institutional withdrawal, or another clearly defined event.
For each case, identify the event time and confirm that the individual was part of the source cohort and at risk immediately before the event.
Incident cases should be distinguished from prevalent cases already experiencing the condition at cohort entry. Including prevalent cases without appropriate design justification can undermine temporal interpretation and alter the population being studied.
Outcome definitions must also be sufficiently consistent across the observation period. If diagnostic criteria or registry coding changed, changes in registry definitions may affect case ascertainment and require harmonization or period-specific analysis.
What Is a Risk Set?
A risk set is the group of cohort members who are under observation and eligible to experience the outcome at a particular event time.
When a case occurs, the researcher identifies other cohort members who are still at risk at that moment. Controls are then sampled from this eligible group according to the prespecified sampling procedure.
For a first-event outcome, individuals who have already experienced the event would ordinarily no longer be eligible as controls for later first-event risk sets. Individuals who have left observation before the case's event time would also be excluded from that risk set.
Entry times matter. A person who has not yet entered the cohort at a case's event time should not be treated as though they had already contributed eligible person-time.
This approach is commonly called risk-set sampling or incidence density sampling.
Why Must Controls Be Selected at the Time Each Case Occurs?
Control selection should reflect the population that could have produced the case at the relevant event time.
If researchers instead select only individuals who remain outcome-free at the end of the entire study, they may exclude people who were eligible controls earlier but subsequently developed the outcome.
Such a strategy can distort the comparison and may introduce bias, particularly when exposure and follow-up time are related.
Risk-set sampling avoids requiring controls to survive outcome-free through the end of follow-up. Under appropriate sampling and analysis conditions, it supports estimation of the exposure-related incidence rate ratio or hazard ratio associated with the underlying cohort process.
This interpretation does not require the outcome to be rare when incidence density sampling is implemented and analyzed appropriately.
Can a Control Later Become a Case?
Yes. A participant selected as a control at one event time may subsequently experience the outcome and become a case.
This is not a design error. At the earlier event time, the participant was legitimately at risk and therefore eligible for control selection.
For example, a patient selected as a control in March may develop the outcome in September. The patient can contribute as a control in the March risk set and as a case in September.
Depending on the sampling scheme, a participant may also be selected as a control for more than one case. The sampling and analysis procedures must accommodate these possibilities.
Removing such participants merely to create mutually exclusive case and control groups can undermine the logic of incidence density sampling.
How Many Controls Should Be Selected per Case?
There is no universally optimal number. The appropriate ratio depends on the available risk sets, exposure prevalence, measurement costs, desired precision, and analytical plan.
One to four controls per case is a familiar practical range, but it is not a methodological requirement. Additional controls can improve precision, with diminishing gains as the number increases.
When exposure assessment is expensive, the marginal value of additional controls should be weighed against their cost. When records are inexpensive and risk sets are sufficiently large, a higher ratio may be feasible.
Some cases may have fewer eligible controls than others, particularly near the end of follow-up or within narrowly defined matching strata.
Study size planning should consider the expected number of incident cases, exposure distribution, within-set matching, and plausible effect sizes rather than choosing a ratio solely because it is conventional.
Should Controls Be Matched to Cases?
Matching may improve efficiency or address aspects of the sampling design when researchers select controls with particular characteristics, such as age, calendar time, institution, or duration of observation.
However, matching is not automatically necessary, and it does not eliminate confounding by itself.
Matching on variables strongly associated with exposure but not sufficiently relevant to the outcome may reduce efficiency. Matching on inappropriate variables can also complicate interpretation or introduce bias.
Risk-set sampling already aligns cases and controls on event time through the sampling process. Additional matching factors should be chosen for explicit methodological reasons.
If individual matching is used, the analysis must account for the matched sets. Ignoring the matching structure may produce inappropriate estimates or standard errors.
When Must Exposure Information Be Measured?
Exposure assessment should correspond to a scientifically meaningful period before the case's event time.
For each matched risk set, researchers should establish a common index date based on the case's event time and evaluate exposure information for cases and controls relative to that date.
For example, a study of medication exposure and subsequent hospitalization might define exposure during a prespecified period preceding the index date.
Using exposure information collected after the index date can introduce temporal ambiguity or bias, particularly if the outcome influences subsequent treatment or measurement.
Registries may contain measurements at irregular intervals. Researchers should therefore investigate whether exposure availability and measurement frequency differ between cases and controls.
When costly exposure measurement is performed after selecting cases and controls, stored specimens or historical records must still represent the relevant pre-event exposure period.
What If Important Variables Were Added to the Registry Later?
Nested case-control designs require comparable exposure and covariate information for selected cases and controls.
If a variable was introduced halfway through registry operation, earlier risk sets may lack the information necessary for the proposed analysis.
Researchers should investigate variables added or removed during registry operation before defining eligible calendar periods or selecting controls.
Restricting analysis to periods with adequate measurements may be reasonable, but it changes the source population and the period to which the findings apply.
Missing measurements should not automatically be treated as absence of exposure.
How Is a Nested Case-Control Study Analyzed?
For individually matched risk sets, conditional logistic regression is a commonly used analytical approach. It compares exposure information within case-specific matched sets while accounting for the sampling structure.
With appropriate incidence density sampling and assumptions, the resulting odds ratio estimates the corresponding incidence rate ratio or hazard ratio parameter from the underlying cohort model.
Unlike a conventional cumulative case-control comparison, this interpretation does not depend on a rare-outcome approximation.
The model should incorporate relevant confounders that can be estimated within the sampling design. Matching variables that are constant within each matched set are ordinarily accounted for through the conditioning rather than estimated as separate within-set effects.
Alternative sampling schemes may require different estimators, weighting, or variance procedures. Analysts should not assume that every dataset labeled nested case-control can be analyzed using an ordinary unmatched logistic regression.
Repeated selection of individuals and complex sampling arrangements may also require appropriate variance estimation.
How Does a Nested Case-Control Study Differ From a Conventional Case-Control Study?
Both designs compare cases with controls, but a nested case-control study defines these groups within an identifiable source cohort and typically samples controls from event-specific risk sets.
A conventional case-control study may also have a well-defined source population, but it does not necessarily originate from a previously enumerated cohort with recorded individual follow-up.
The nested design can provide clearer information about eligibility, event timing, and the population from which cases arose. Its validity nevertheless depends on the quality of those records.
How Does It Differ From a Case-Cohort Study?
| Feature |
Nested Case-Control |
Case-Cohort |
| Comparison group |
Controls sampled from the risk set at each case's event time |
A subcohort sampled from the source cohort, commonly at baseline |
| Sampling structure |
Case-specific control sets |
A common sampled subcohort used across eligible outcome analyses |
| Potential reuse |
Individuals may serve as controls repeatedly and later become cases |
Subcohort members can contribute to analyses of multiple outcomes |
| Typical analysis |
Conditional logistic regression for matched risk sets |
Specialized weighted survival or risk-set methods |
| Practical advantage |
Exposure assessment concentrated on cases and sampled risk-set controls |
One sampled subcohort may support several outcome investigations |
Neither design is universally superior. The choice depends on the research questions, outcomes, available cohort data, and cost of exposure measurement.
When Is a Full Cohort Analysis Preferable?
If exposure, covariate, and outcome information are already available for every cohort member at minimal additional cost, analyzing the full cohort may be more efficient and straightforward.
A nested case-control design is especially attractive when obtaining detailed exposure information for everyone would be expensive or impractical.
For example, a registry may contain basic information on 100,000 participants but require manual review of thousands of clinical notes to classify a specialized exposure. Sampling controls can substantially reduce that workload.
However, the nested design sacrifices some information relative to an otherwise comparable full-cohort analysis. Researchers should justify the efficiency benefit rather than selecting the design simply because the outcome is uncommon.
What Biases Should You Consider?
A nested case-control design does not automatically eliminate the limitations of the source registry.
Selection into the registry may affect the population represented by the cohort. Researchers should examine systematic differences between registry participants and nonparticipants when broader inference is intended.
Outcome misclassification, exposure measurement error, missing covariates, and unmeasured confounding can also threaten validity.
Incorrect risk-set construction introduces additional problems. Controls selected from individuals who were not yet enrolled, had already experienced the event, or were no longer under observation may not represent the appropriate person-time at risk.
Another concern arises when follow-up ends selectively. If censoring is related to exposure and future outcome risk, analyses may require additional assumptions or methods to address informative censoring.
The design's statistical efficiency should not be mistaken for protection against these problems.
What Should Be Documented Before Analysis?
A defensible protocol should specify the source cohort, eligibility criteria, cohort entry, outcome definition, follow-up rules, case ascertainment, risk-set construction, control sampling, matching factors, exposure windows, and analysis method.
Researchers should also document whether controls are sampled with replacement across risk sets, whether future cases can serve as controls, and how repeated selection is handled.
The STROBE case-control reporting checklist calls for transparent descriptions of case ascertainment, control selection, matching, data sources, potential bias, and statistical methods.
These details are essential for evaluating whether the analysis preserves the intended cohort-based sampling structure.
