Case-Cohort Study

Case-cohort study is an efficient two-phase epidemiological sampling architecture introduced by biostatistician Ross L. Prentice in 1986 to evaluate expensive molecular, genomic, proteomic, or environmental exposure assays within large prospective cohort studies. In a full prospective cohort of size N (frequently 50,000 to 500,000 participants in modern biobanks), baseline biological specimens—such as cryopreserved serum, plasma, buffy-coat leukocytes, or urine—are collected and archived for all participants at enrollment (t = 0), and the entire cohort is followed longitudinally for incident disease endpoints. Rather than thawing and assaying specimens for all N participants, a case-cohort design selects two components: (1) a random subcohort of size n_sub drawn via simple or stratified random sampling from the baseline cohort without regard to future disease status, and (2) all incident cases of the target disease who develop the outcome during follow-up outside the random subcohort. Because the random subcohort is a direct representative probability sample of the baseline source cohort and its person-time at risk, a single subcohort can simultaneously serve as the reference comparison group for multiple distinct disease endpoints while permitting direct estimation of both cumulative risk ratios (RR) and Cox proportional hazard ratios (HR) without invoking the rare-disease assumption.

Overcoming the Cost Barrier of Biobank Molecular Assays

In chronic disease epidemiology and public health research, prospective cohort studies eliminate retrospective recall bias and reverse causality because biomarkers are collected years before clinical diagnosis. However, when only 1% to 3% of cohort members develop an incident malignancy, stroke, or neurodegenerative disorder over a 10-year follow-up period, more than 97% of the laboratory budget in a full-cohort assay is spent measuring non-cases. High-resolution assays—such as whole-genome bisulfite methylation sequencing, untargeted liquid chromatography–mass spectrometry (LC-MS) metabolomics, multiplex cytokine panels, or persistent organic pollutant quantification—often cost hundreds of dollars per specimen and consume irreplaceable biobank aliquots.

By restricting laboratory assays to the random subcohort (n_sub, typically 3% to 10% of the full cohort N, or roughly 1 to 4 times the expected number of incident cases) plus the supplementary incident cases outside the subcohort, the case-cohort design captures 80% to 95% of the statistical precision of the full N-person cohort while reducing analytical assay costs and specimen depletion by 85% to 95%.

Overlap Between the Subcohort and Incident Cases

A defining structural feature of the case-cohort design is that the random subcohort (S) is selected at baseline (t = 0) completely independent of future clinical outcomes. Consequently, if the overall cumulative incidence of the disease in the full cohort is p = D / N and the subcohort sampling fraction is α = n_sub / N, then approximately α × D members of the random subcohort will naturally develop the incident disease during longitudinal follow-up. These subcohort cases (S ∩ D) play a dual role in the statistical analysis:

• Prior to their event diagnosis time t_j, they contribute baseline person-time at risk to the subcohort denominator just like disease-free subcohort members.
• At their event diagnosis time t_j, they serve as incident cases in the numerator alongside the non-subcohort cases (D \ S) sampled from the remainder of the cohort.

Direct Estimation of Risk Ratios and Odds Ratios Without Rare-Disease Approximations

Let a full cohort of N baseline-free individuals experience D incident cases (D₁ exposed, D₀ unexposed) over a fixed follow-up interval. Suppose a random subcohort of size n_sub contains m₁ exposed individuals and m₀ unexposed individuals at baseline (including those who subsequently become cases). Because the subcohort is a simple random sample of the entire baseline cohort (m₁ / m₀ estimates the baseline exposure ratio N₁ / N₀ in the entire cohort, not the non-cases), the cross-product ratio of exposed-to-unexposed cases divided by exposed-to-unexposed subcohort members directly estimates the cumulative incidence ratio (Risk Ratio, RR):

RR_case-cohort = ( D₁ / D₀ ) / ( m₁ / m₀ ) ≈ ( D₁ / N₁ ) / ( D₀ / N₀ )

This exact risk-ratio property holds regardless of how common the disease is in the cohort—no rare-disease approximation is required.

Weighted Cox Proportional Hazards and Pseudo-Likelihood Estimators

To analyze right-censored time-to-event data under David Cox's proportional hazards model, h(t | X_i) = h₀(t) exp(β^TX_i), the full-cohort partial likelihood sums across all cohort members at risk at each ordered event time t_j. In a case-cohort study, covariates X_i are observed only for subcohort members and non-subcohort cases. Epidemiologists and biostatistics researchers use three principal weighting schemes for the risk-set denominator at event time t_j:

1. Prentice (1986) Unweighted Pseudo-Likelihood: At event time t_j, the risk set S̃(t_j) consists of all members of the random subcohort who remain at risk at t_j, plus the failing case j itself if case j arose outside the subcohort (w_i = 1 for all members of S̃(t_j)). Non-subcohort cases do not enter the risk sets of earlier failure times t_k < t_j.
2. Self and Prentice (1988) Estimator: Risk sets consist solely of random subcohort members at risk at t_j; non-subcohort cases appear only in the numerator at their own failure time t_j and receive weight 0 in the denominator.
3. Barlow (1994) and Borgan et al. (2000) Inverse-Probability Weighted (Horvitz–Thompson) Estimator: Weights every disease-free subcohort member by the inverse of the subcohort sampling fraction, w_i = 1 / α = N / n_sub (or stratum-specific inverse probabilities w_i,s = N_s / n_sub,s in a stratified case-cohort design), while giving all cases weight w_i = 1 at their failure time, reconstructing the exact total person-time of the full cohort N.

Because non-subcohort cases enter the pseudo-likelihood and because subcohort members appear across multiple risk sets, standard Fisher information matrices underestimate standard errors. Valid confidence intervals require the Self–Prentice asymptotic covariance adjustment or Barlow's jackknife / robust sandwich variance estimator (coxph with cluster/robust weights or cch in the R survival package).

When a baseline confounder or exposure stratum is rare in the cohort (for example, an occupational exposure or minority ancestry group), stratified case-cohort sampling (Borgan et al., 2000) oversamples informative baseline strata into the subcohort and applies stratum-specific inverse-probability weights w_s = N_s / n_sub,s, boosting statistical power while preserving unbiased hazard ratio estimation.

A practical laboratory consideration in case-cohort studies is assay batch drift and differential specimen freeze-thaw history. If the random subcohort is assayed at the beginning of a study and incident cases are assayed years later as they accrue, systematic laboratory drift produces differential exposure misclassification. Best practice requires storing subcohort aliquots under identical liquid-nitrogen conditions and plating subcohort specimens and accrued case specimens together in randomized, blinded laboratory batches, while applying repeat-biomarker subsamples to correct for regression dilution bias.

When is a nested case-control study preferred over a case-cohort study?

Although the case-cohort design is superior when investigators plan to study multiple disease endpoints (such as myocardial infarction, ischemic stroke, type 2 diabetes, and colorectal cancer) using the same baseline biomarker panel, a nested case-control study is preferred when the biomarker degrades slowly over years of freezer storage, varies strongly with seasonal blood-draw date, or requires tight matching on exact follow-up duration t_j and analytic laboratory batch.

Why must subcohort cases not be excluded from the baseline random subcohort?

Excluding individuals who later develop the disease from the baseline subcohort transforms the comparison group from a representative sample of the total baseline population (N₁ / N₀) into a sample of survivors/non-cases (b / d). Doing so destroys the case-cohort design's ability to directly estimate the cumulative Risk Ratio (RR) and biases the Cox proportional hazards risk sets at all event times prior to each subcohort case's diagnosis.

1. Prentice, R. L. (1986). A case-cohort design for epidemiologic cohort studies and disease prevention trials. Biometrika, 73(1), 1–11. https://doi.org/10.1093/biomet/73.1.1
2. Self, S. G., & Prentice, R. L. (1988). Asymptotic distribution theory and efficiency results for case-cohort studies. The Annals of Statistics, 16(1), 64–81. https://doi.org/10.1214/aos/1176350691
3. Barlow, W. E. (1994). Robust variance estimation for the case-cohort design. Biometrics, 50(4), 1064–1072. https://doi.org/10.2307/2533444
4. Borgan, Ø., Langholz, B., Samuelsen, S. O., Goldstein, L., & Pogoda, J. (2000). Exposure stratified case-cohort designs. Lifetime Data Analysis, 6(1), 39–58. https://doi.org/10.1023/A:1009661900674