When Is a Case-Cohort Design Useful?

What If You Want to Study Several Outcomes Without Measuring Everyone in a Large Cohort?

Suppose a research institution has followed 50,000 adults for several years and stored biological specimens at baseline. Researchers want to investigate whether several biomarkers are associated with cardiovascular disease, cancer, and other outcomes.

Analyzing every specimen would be expensive. A nested case-control study could reduce costs by selecting controls for each outcome, but different outcomes might require different sampled control sets.

Could researchers instead select one representative subset of the cohort, measure its biomarkers, and reuse that information across several outcome analyses?

This is the central appeal of a case-cohort design.

The approach can be efficient when detailed exposure measurements are expensive and several outcomes are of interest. However, it introduces a sampling structure that must be reflected in the analysis, and its advantages depend on the underlying cohort and research objectives.

Use a Case-Cohort Design When One Subcohort Can Support Several Outcome Analyses

How Case-Cohort Sampling Works and When It Is Efficient

What Is a Case-Cohort Study?

A case-cohort study is an efficient sampling design conducted within a defined cohort.

Rather than measuring an expensive exposure for every cohort member, researchers select a subcohort using a prespecified sampling procedure. They then include eligible cases of the outcome occurring in the parent cohort, including cases outside the sampled subcohort.

In a common implementation, the subcohort is randomly sampled at baseline independently of subsequent outcome occurrence.

Because subcohort membership is not determined by whether a person later becomes a case, the subcohort can contain individuals who subsequently experience the outcome.

The subcohort provides sampled information about the parent cohort's population experience, while the additional cases provide information about outcome events not already represented among sampled subcohort members.

Appropriate statistical methods account for the resulting unequal sampling and risk-set representation.

How Does the Subcohort Differ From an Ordinary Control Group?

In a conventional case-control study, controls are sampled according to a strategy tied to the case-generating population and outcome definition.

In a case-cohort design, the subcohort is generally selected independently of future outcome status from the parent cohort.

Consequently, the subcohort is not simply a group of people who never experience the disease.

Some subcohort members may become cases during follow-up. They remain members of the sampled subcohort, and their case events are incorporated according to the analytical method.

This feature is essential to the design. Excluding future cases from the subcohort would change the sampling mechanism and could undermine its representativeness.

The principle that controls and sampled population experience must correspond to the case-generating population remains relevant, but the operational sampling approach differs from conventional noncase controls.

How Is a Case-Cohort Design Different From a Nested Case-Control Design?

Both designs reduce the need to measure expensive exposures throughout a large cohort. Their main distinction is how the comparison sample is selected.

Neither design is universally more efficient. Their relative performance depends on outcome frequency, exposure variation, sampling fractions, measurement costs, and the intended analyses.

The comparison with nested case-control sampling is particularly important when deciding whether the same expensive exposure measurements will be needed for several outcomes.

Why Is Case-Cohort Sampling Useful for Multiple Outcomes?

A major advantage is that the subcohort is selected independently of the outcomes that researchers subsequently investigate.

Suppose a cohort contains stored blood samples and investigators want to study three diseases. They could assay the selected subcohort once and use its measurements in analyses of all three outcomes.

Cases of each disease occurring outside the subcohort would receive the relevant exposure measurements as needed.

This can reduce duplicated laboratory work compared with drawing entirely separate comparison samples for each outcome.

However, the efficiency advantage depends on overlap among cases, assay requirements, and the number of outcomes.

If only one outcome is being investigated, a nested case-control design may be equally or more efficient under particular conditions.

If all exposure measurements already exist for the entire cohort, neither sampled design may provide a meaningful cost advantage.

How Is the Subcohort Selected?

In a standard case-cohort design, researchers randomly select a subcohort from the eligible parent cohort at a defined sampling point, commonly baseline.

Simple random sampling is one option. Stratified sampling may also be used when researchers want sufficient representation of particular groups.

For example, a cohort might be stratified by recruitment center or another prespecified characteristic before sampling.

When sampling probabilities differ across strata, the analysis must account for those differences.

The subcohort should not be selected according to future disease status, because that would undermine the intended outcome-independent sampling structure.

Likewise, selecting only participants who remain disease-free at the end of follow-up would not reproduce the standard case-cohort design.

Can a Subcohort Member Also Be a Case?

Yes. Subcohort membership and case status are not mutually exclusive.

Suppose a subcohort contains 1,000 people sampled at baseline. During follow-up, 30 of those individuals develop the outcome.

Those 30 individuals remain members of the subcohort and also contribute case events.

Cases arising outside the subcohort are included according to the study protocol.

The statistical analysis must handle the overlapping roles correctly. Simply counting subcohort cases twice as independent participants would be inappropriate.

This distinction is particularly important when preparing datasets for specialized case-cohort regression procedures.

Why Can Case-Cohort Sampling Reduce Costs?

The main efficiency gain occurs when detailed exposure assessment is expensive but outcome ascertainment is available for the full cohort.

For example, a cohort may already have information about diagnoses, follow-up dates, and baseline characteristics, while laboratory assays for stored specimens remain costly.

Researchers can measure biomarkers in the sampled subcohort and relevant cases instead of assaying every specimen.

Cost savings depend on the subcohort sampling fraction, the number of additional cases, and the extent to which measurements can be reused across outcomes.

Additional costs may arise from data management, specialized statistical analysis, and laboratory procedures.

Researchers should compare the total expected cost of case-cohort sampling with full-cohort and nested case-control alternatives rather than assuming that sampling automatically produces the greatest savings.

How Large Should the Subcohort Be?

There is no universal optimal subcohort fraction.

The required size depends on the number of cases, expected exposure distribution, desired precision, sampling design, and analytical objectives.

A larger subcohort generally provides more information about the parent cohort's exposure distribution and person-time experience, but it also increases measurement costs.

A very small subcohort may contain too few exposed participants or insufficient representation of important strata.

When several outcomes are investigated, subcohort-size planning should consider the requirements of each intended analysis rather than only the most common outcome.

Researchers may need simulation-based study-size assessments when sampling is stratified, exposures are uncommon, or several outcomes have different event patterns.

How Are Case-Cohort Data Analyzed?

Case-cohort data require methods that account for the subcohort sampling mechanism and the inclusion of additional cases.

Weighted or modified Cox proportional hazards methods are commonly used for time-to-event outcomes.

Prentice's original approach and subsequent methods developed by other researchers provide ways to estimate regression parameters using sampled cohort data.

The appropriate method depends on the sampling design, whether sampling is stratified, the risk-set representation, and the intended estimand.

Variance estimation also requires special attention because the subcohort is sampled and observations may contribute through overlapping roles.

Robust or design-appropriate variance estimators are commonly used, but their implementation must correspond to the selected estimation method.

Researchers should not simply combine all cases and subcohort members into an ordinary unweighted Cox regression without considering the case-cohort sampling structure.

Can Case-Cohort Studies Estimate Hazard Ratios?

Yes. Under appropriate assumptions and analysis, case-cohort designs can estimate hazard ratios associated with exposures.

A hazard ratio compares instantaneous event rates between groups conditional on the modeled covariates and time structure.

It is not automatically a risk ratio or a statement about absolute disease probability.

The proportional hazards assumption may be relevant when a Cox model is used. If the exposure association changes over time, alternative specifications may be required.

Researchers should report the estimation method, sampling weights where applicable, variance procedure, and relevant model assumptions.

They should also distinguish association estimates from causal effects, which require additional assumptions about confounding, selection, and measurement.

Can Case-Cohort Designs Support Absolute-Risk Estimation?

Potentially, when the necessary cohort information, sampling probabilities, and appropriate estimation methods are available.

The parent cohort may provide outcome counts, person-time, and population denominators.

Subcohort sampling can also support estimation of exposure distributions or model parameters needed for particular absolute-risk procedures.

However, absolute risk is not obtained simply by dividing the number of cases by the number of subcohort participants.

That calculation would ignore the sampling structure and the relationship between the subcohort and the full cohort.

The distinction between relative associations and absolute risk or prevalence estimation remains important.

What Happens When Exposure Changes Over Time?

Time-varying exposures complicate case-cohort analysis because measurements must represent the relevant exposure history during follow-up.

If only baseline exposure is available, researchers may be limited to questions about baseline status or must make assumptions about exposure stability.

Repeated exposure measurements can support more detailed analyses, but the sampling and estimation methods must accommodate their timing.

For example, medication use may change repeatedly during follow-up. A baseline prescription record may not represent exposure several years later.

Researchers should ensure that exposure information used in an analysis precedes the relevant outcome event and corresponds to the intended risk period.

These requirements can influence whether a case-cohort design is preferable to a full-cohort analysis or another sampling approach.

Can Stratified Case-Cohort Sampling Improve Efficiency?

Yes. Researchers may sample the subcohort at different rates across predefined strata when doing so improves representation or statistical efficiency.

For example, an uncommon exposure group might be oversampled to ensure sufficient information for estimating an exposure association.

However, unequal sampling probabilities must be incorporated into the analysis.

Otherwise, the sampled exposure distribution may not represent the parent cohort in the way required for valid estimation.

Stratification should be planned before outcome-associated exposure patterns are examined and supported by appropriate study-size calculations.

Specialized weighting and variance methods may be necessary.

When Is a Full-Cohort Analysis Preferable?

Full-cohort analysis may be preferable when exposure information is already available for all cohort members at little additional cost.

It may also be attractive when researchers need highly flexible time-varying analyses, complex prediction models, or detailed estimation of several population quantities.

Sampling can reduce information and introduce analytical complexity without providing a meaningful benefit when no expensive measurements are avoided.

Conversely, case-cohort sampling becomes more attractive when the full cohort is very large and exposure assessment requires costly laboratory or manual procedures.

The decision should compare the information gained and costs incurred under each feasible design.

What Are the Main Limitations of Case-Cohort Studies?

Case-cohort designs require a suitable parent cohort with reliable eligibility, follow-up, and outcome ascertainment.

The sampled subcohort must be selected through a defensible probability mechanism, and relevant sampling information must be retained.

Specialized analysis may be more demanding than ordinary full-cohort regression, particularly with stratified sampling, time-varying exposures, or multiple outcomes.

Missing specimens, incomplete exposure records, and differential follow-up can also threaten validity.

Finally, the design may be inefficient if the subcohort is too small for important exposure comparisons or if the number of cases outside the subcohort is unexpectedly large.

These limitations do not make case-cohort research inherently weaker. They identify the conditions under which its efficiency advantages must be evaluated carefully.

Studying Three Diseases Using One Sampled Subcohort

For simplicity, assume these outcomes occur in distinct individuals and that 15, 22, and 13 of the respective cases belong to the sampled subcohort.

The researchers measure the relevant biomarkers in the 2,000 subcohort members and the 950 additional cases outside the subcohort.

Under the stated assumptions, this requires 2,950 unique specimens rather than 40,000.

The difference is substantial, but the calculation excludes laboratory setup, storage, data management, and specialized analytical costs.

The researchers should not treat the 2,000 subcohort members as permanently disease-free controls. Some of them became cases, and that overlap is part of the intended design.

Nor should they calculate disease incidence by dividing the 1,000 cases by the 2,950 measured participants.

The full cohort provides the relevant outcome and follow-up information, while the sampled subcohort supports efficient exposure analysis.

This is the central advantage of case-cohort sampling: a single probability-based comparison sample can contribute to several outcome analyses without requiring expensive measurements for the entire cohort.

Common Misconceptions About Case-Cohort Designs

How Should You Decide Whether Case-Cohort Sampling Fits Your Research?

Start by determining whether the parent cohort contains the outcome and follow-up information required for the intended analyses.

Then assess whether expensive exposure measurements can be concentrated on a sampled subcohort and relevant cases without compromising the scientific objectives.

The protocol should specify how the subcohort is sampled, how cases are identified, which exposure measurements are required, and how overlap between subcohort membership and case status will be handled.

Researchers should also document the estimation method, sampling weights where applicable, and variance procedure.

The strongest justification for case-cohort sampling is its ability to answer the intended questions efficiently while preserving valid inference from the underlying cohort.

Before Choosing a Case-Cohort Design, Verify the Sampling and Analysis Plan

Frequently Asked Questions About Case-Cohort Studies

Case-Cohort Sampling Is Most Useful When One Subcohort Can Serve Several Analyses

How to Cite This Guide

This guide is intended to be read, shared, and used in research, teaching, and academic work. If you draw on its ideas, explanations, or other content, please acknowledge the source by citing the guide. Doing so gives appropriate credit and helps your readers locate the original resource.