Audit a case-cohort sample before modelling

Identify the sampled components

A case-cohort study combines a sampled subcohort with cases arising in the source cohort. The subcohort is selected independently of future case status under the specified sampling design; some subcohort members may later become cases. Treating the assembled records as an ordinary cohort can misrepresent the risk sets and uncertainty.

The survival package case-cohort documentation describes estimators that explicitly use subcohort membership and original cohort sizes. Before fitting any model, obtain the sampling protocol, selection list, sampling strata and the full cohort counts used to construct the sample.

Reconcile membership without duplicating people

Create separate indicators for subcohort membership and case status. Keep a stable person identifier and verify that a case already in the subcohort is not duplicated as an additional independent person. Record entry, follow-up and event definitions consistently for both sampled components.

For stratified sampling, reconcile the number selected with the original population in every stratum. Check whether the selection probabilities differ and whether the estimator supports that design. Missing biomarker measurements can create another selection stage; preserve its reasons rather than silently dropping records and retaining the original sampling interpretation.

Hypothetical example

A fictional cohort contains 5,000 participants. Investigators select a subcohort of 500 and identify 200 cases during follow-up, 20 of whom are already in the subcohort. The assembled set contains 500 + 200 − 20 = 680 unique people, not 700.

The 680 records are not a simple random sample of the cohort. Case inclusion and subcohort selection have different roles. The arithmetic checks record reconciliation only; it does not specify the analysis weights or variance estimator.

Match software to the sampling design

Document the chosen case-cohort estimator, handling of strata and uncertainty calculation. Check the implementation's restrictions on missing values and records outside the subcohort. Validate the data structure with a small synthetic cohort whose membership can be traced manually.

Report source-cohort size, subcohort selection, case overlap, exclusions and the method used. Do not present a naive standard error simply because it is the first value printed. Use the table grain and primary-key worksheet to preserve person identity, then add explicit sampling indicators and population counts. If those design records cannot be reconstructed, resolve that uncertainty before claiming a design-consistent analysis.

Source verification: 6 November 2026. Examples are hypothetical.