Case-Cohort Studies: Design and Efficiency

A case-cohort study picks its comparison group differently from almost every other design built inside a cohort: it draws one random subcohort, once, at or near baseline — before anyone knows who will develop any outcome — and then reuses that same subcohort as the comparison group for every outcome studied from that cohort. That single choice is what separates a case-cohort design from a nested case-control study, whose comparison group is instead resampled from scratch, per case, at the moment each case occurs. This guide stays on that specific mechanical difference: how the subcohort is built, why it can serve multiple outcomes at once, what that reuse costs in the analysis stage, and a reproducible worked example showing the efficiency gain when a cohort is used to study more than one endpoint.

What a case-cohort design actually does

Start with a cohort already enrolled and under follow-up for one or more outcomes — the same starting point a cohort study uses. Instead of measuring an expensive exposure (a stored-serum assay, a genotype call, a manual chart abstraction) on every enrolled member, a case-cohort design selects a subcohort: a simple random sample (sometimes a stratified random sample, to guarantee representation across sites or strata) of the full cohort, typically 5–20% of it, drawn at or shortly after baseline. Exposure is measured on that subcohort. It is also measured on every person in the full cohort who goes on to become a case of the outcome under study, whether or not that person happened to land in the subcohort. The subcohort is the design’s reference group: cases are compared against it, weighted appropriately (see below), to estimate the exposure–outcome association.

The method traces to Ross Prentice’s 1986 paper in Biometrika, which introduced the case-cohort design specifically to make expensive-exposure epidemiologic cohort studies and disease-prevention trials affordable without discarding the cohort’s prospective structure. The core idea hasn’t changed since: sample the comparison group once, from the whole cohort, independent of case status or event timing, and get a reusable reference set out of it.

Case-cohort vs. nested case-control: where the sampling actually diverges

Both designs exist to solve the same underlying problem — measuring an expensive exposure on an entire cohort is often infeasible, so you measure it on cases plus a smaller, chosen comparison group instead. Where they diverge is entirely in how that comparison group gets built:

Nested case-control uses risk-set (incidence-density) sampling: every time a case occurs, controls are drawn from the risk set at that exact moment — the cohort members still under follow-up and still outcome-free at that instant. The comparison group is assembled case-by-case, continuously, throughout follow-up, and is specific to one outcome’s timing.
Case-cohort draws the subcohort once, at baseline, with no reference to when (or whether) any case occurs. The same subcohort stands as the comparison group for every case, for every outcome, for the entire follow-up period.

That difference has a direct consequence for studies with more than one outcome. A nested case-control study answering two separate research questions from the same cohort needs two separate risk-set-sampling exercises — a fresh round of control selection (and, if the exposure isn’t already banked, fresh assays) for each outcome, because each outcome has its own case timing and therefore its own risk sets. A case-cohort study answering the same two questions reuses the identical subcohort for both: the subcohort was never tied to a specific outcome’s timing in the first place, so nothing about adding a second endpoint requires resampling it.

The efficiency payoff: multiple endpoints from one subcohort

This is the case-cohort design’s strongest practical argument, and it’s easiest to see with numbers rather than description. The illustration below is a reproducible seeded simulation — not a real published study or dataset — built to make the mechanics concrete; every figure in it was computed by a deterministic script (fixed seed 20260826, Mulberry32 PRNG), not estimated or invented.

Take a cohort of 10,000 participants under follow-up for two independent outcomes: Outcome A with 2% cumulative incidence, Outcome B with 1.5% cumulative incidence. The simulation produced 190 cases of A and 153 cases of B.

Two separate nested case-control studies (1 case : 4 controls, run independently per outcome, since each outcome’s risk sets are distinct): 190 × 5 = 950 assays for Outcome A, plus 153 × 5 = 765 assays for Outcome B — 1,715 assays total, or 17.2% of the full cohort.
One case-cohort design (10% random subcohort, drawn once): the subcohort came out to 1,007 people. Of the 190 Outcome A cases, 173 fell outside the subcohort and needed a fresh assay (the other 17 were already covered as subcohort members). Of the 153 Outcome B cases, 136 fell outside the subcohort. Total: 1,007 + 173 + 136 = 1,316 assays, or 13.2% of the full cohort — roughly 77% of what the two separate nested case-control studies required, for the same two endpoints, because the 1,007-person subcohort assay cost was paid once and then reused rather than paid twice.

The gap widens with every additional endpoint studied from the same cohort: a third or fourth outcome adds only its own out-of-subcohort cases to the case-cohort total, while each additional outcome under nested case-control requires an entirely new round of risk-set sampling and control assays from zero. This is the specific efficiency case-cohort makes over nested case-control — not that it is cheaper for a single outcome (the two designs are broadly comparable there), but that the fixed cost of building the comparison group is paid once and amortized across however many outcomes the cohort eventually gets used to study, which matters most for biobank-style cohorts assembled for open-ended future use rather than one prespecified question.

The price of reuse: why case-cohort needs a weighted analysis

Reusing one subcohort as the comparison group for multiple, differently-timed outcomes has a real analytical cost. A subcohort member represents more than just themselves in the risk set at any given event time — unlike risk-set sampling, where the sampling fraction naturally tracks person-time (see nested case-control studies), the case-cohort subcohort’s fixed, baseline-drawn structure means an ordinary (unweighted) Cox or logistic model applied directly to subcohort-plus-cases data produces biased estimates. Case-cohort analysis instead requires a weighted estimator built specifically for this sampling structure:

Prentice’s original (1986) pseudolikelihood and the Self & Prentice (1988) modified pseudolikelihood were the first estimators built for this design, each handling the subcohort’s role in the risk set slightly differently.
Barlow-style weighting (1994) extends the idea into a weighted Cox proportional-hazards model: subcohort members are given a weight related to the inverse of the subcohort’s sampling fraction, cases outside the subcohort are added in at their observed event times, and the model is fit on that weighted risk set.
Because the same subcohort member can appear in the risk set at many different event times across follow-up, the resulting estimates aren’t independent across those appearances the way a nested case-control’s matched sets are — case-cohort analyses need a robust (“sandwich”) variance estimator to get correct standard errors, not the model’s default variance.

None of this is optional bookkeeping: fitting an off-the-shelf logistic or Cox model to case-cohort data without one of these weighted approaches is a common and consequential analysis error, not a simplification that happens to work.

Choosing between case-cohort and nested case-control

Both designs solve the same expensive-exposure problem; the right one depends on what the cohort is for:

Favor case-cohort when the cohort will (or might) be used to study multiple outcomes — including outcomes not yet defined at baseline, which is common for banked-specimen cohorts and biobanks — since the subcohort’s value compounds with each additional endpoint.
Favor nested case-control when there’s a single, clearly defined outcome, especially one where the rare-disease assumption would otherwise be a concern: risk-set sampling lets the resulting odds ratio estimate the incidence rate ratio directly, and the matched conditional-logistic analysis it supports is simpler to specify and explain than a weighted case-cohort model.
Case-cohort’s subcohort size is usually fixed as a proportion of the whole cohort rather than tuned to a controls-per-case ratio for one specific outcome, so per-outcome statistical precision is somewhat less finely adjustable than it is in a nested case-control study built around that one outcome.

Common pitfalls

Treating subcohort cases as double-counted or excluded. A subcohort member who later becomes a case is not removed from the subcohort and is not counted twice for exposure measurement — their exposure was already captured as a subcohort member, and the weighted analysis accounts for their dual role (comparison-group member up to their event, case at their event) through the weighting scheme itself, not through duplicate records.
Fitting an unweighted model. Plugging subcohort-plus-cases data into ordinary logistic regression or an unweighted Cox model, as if it were a standard case-control dataset, produces biased effect estimates — the fixed, baseline-drawn subcohort doesn’t have the person-time-proportional sampling property that would make that valid.
Using the model’s default standard errors. Because subcohort members contribute to multiple risk sets across follow-up, their contributions are correlated; a robust variance estimator is required, not the default output of a standard Cox or logistic fit.
Confusing case-cohort with nested case-control when quoting design efficiency. The two designs’ comparison-group logic is genuinely different (baseline random sample vs. per-case risk-set sample) — “cheaper than measuring the whole cohort” is true of both, but the multi-outcome reuse advantage above is specific to case-cohort.

Frequently asked questions

Is a case-cohort study the same as a nested case-control study?

No. Both sample a comparison group from a cohort instead of measuring exposure on everyone, but a case-cohort study draws one random subcohort once, at baseline, independent of case timing; a nested case-control study draws controls from the risk set at each case’s specific event time, continuously through follow-up. They require different analysis methods and have different multi-outcome properties.

Can one case-cohort subcohort really be used for more than one outcome?

Yes — that’s the design’s central practical advantage. Because subcohort selection never depended on any outcome’s timing, the same subcohort works as the comparison group for any outcome measured in that cohort, including outcomes defined after the subcohort was drawn.

Why can’t a case-cohort study be analyzed with ordinary logistic regression?

Because the subcohort isn’t sampled with a person-time-proportional structure the way a risk set is. An ordinary model applied directly to subcohort-plus-cases data produces biased estimates; case-cohort data needs a weighted estimator (Prentice, Self-Prentice, or Barlow-style weighting) with a robust variance estimator.

How large should the subcohort be?

There’s no universal figure — it’s a proportion of the full cohort (commonly in the 5–20% range) chosen to give adequate precision for the rarest outcome the subcohort will need to support, balanced against the cost of measuring exposure on that many people. A subcohort sized for one modestly common outcome may still be adequate as a second, rarer outcome is added later, which is part of the design’s appeal for open-ended cohorts.

Does a case-cohort design still work if the outcome is rare?

Yes, and it doesn’t rely on the rare-disease assumption to do so — the weighted case-cohort estimators (Prentice/Self-Prentice/Barlow) target the same hazard-ratio-type estimand a full-cohort survival analysis would, in the same way nested case-control’s risk-set sampling avoids relying on that assumption.

Case-Cohort Studies: Design, Subcohort Reuse, and Efficiency

Ask about Case-Cohort Studies: Design, Subcohort Reuse, and Efficiency