Why Participant Overlap Matters
A validation cohort is informative when it provides evidence from participants who did not contribute to discovery. If a person appears in both cohorts, the two data sets are not fully independent at the participant level. This can happen even when the records have different study IDs, were generated at different sites, or contain different omics measurements. For example, one person may have contributed a tumor specimen to a discovery study and a later specimen to a validation study.
The aim of an overlap check is not merely to find identical rows. It is to identify whether the same person contributed data to both cohorts, to distinguish confirmed matches from uncertain candidates, and to describe how the overlap was handled. The check should be performed before interpreting validation results as independent evidence. A finding of no detected matches is not proof that overlap is impossible: the strength of that conclusion depends on the identifiers and records available for comparison.
This definition is at the person level. An individual may contribute several specimens, visits, assays, or omics layers. Counting those records as separate people inflates cohort sizes and can conceal overlap. Conversely, two records from one person need not represent the same physical specimen. Participant overlap and specimen duplication are related checks, but they answer different questions.
Evidence for a Match
Start by identifying which records can be compared and what each identifier represents. Directly comparable study IDs are useful only if their scope is known: an ID might identify a participant, a specimen, a visit, or a data submission. Recruitment logs and consent or enrollment records can clarify whether apparently different records refer to the same person. Other potentially informative fields include birth year, sex recorded at enrollment, recruitment site, dates, and specimen relationships. These fields should be used only when permitted and appropriate; a similarity in common attributes alone is not definitive identity evidence.
A privacy-conscious approach can use a trusted data custodian to standardize identifiers and generate a consistent, protected participant key for both cohorts. For example, if both data holders apply the same keyed transformation to identically normalized participant identifiers, the resulting keys can be compared without exchanging the original identifiers. A plain hash of a predictable identifier may be vulnerable to guessing, so it should not be treated as secure de-identification. Access, purpose, retention, and any linkage should follow applicable approvals and data-use agreements.
When direct identifiers are unavailable or inconsistent, record linkage may compare several attributes to identify candidate pairs. Such linkage is imperfect: overly permissive rules can create false matches, while overly strict rules can miss genuine matches. A candidate match should therefore be assessed using multiple pieces of evidence and, where permitted, adjudicated against authoritative recruitment records. Keep separate categories for confirmed matches, rejected candidates, and unresolved cases rather than silently converting uncertainty into a yes-or-no result.
Worked Example: Quantifying Confirmed Overlap
Suppose a hypothetical discovery cohort contains 140 distinct participants and a validation cohort contains 90. A shared, protected participant key identifies six validation participants who also appear in discovery. Assume that the key is generated consistently for both cohorts, has been checked against enrollment records, and refers to participants rather than specimens.
The confirmed overlap count is \(k=6\). The validation overlap proportion is \[ \frac{k}{|V|}=\frac{6}{90}=0.0667, \] or \(6.67\%\). Thus, six of the 90 validation participants also contributed to discovery, and 84 validation participants have no confirmed discovery match.
The number of distinct people across the two cohorts is found by subtracting the intersection once: \[ |D\cup V|=|D|+|V|-|D\cap V|=140+90-6=224. \] Adding the cohort sizes without subtracting the six shared people would give 230 records of cohort membership, not 230 distinct people. This arithmetic checks consistency of the counts; it does not establish that the validation analysis is unbiased or that the 84 unmatched participants are representative.
From Candidate Matches to a Defensible Count
A practical assessment separates the matching process into stages. First, deduplicate records within each cohort at the participant level. Then compare the cohorts using available identifiers, preserving the mapping from each candidate match to the evidence that generated it. Resolve candidates where possible, and report confirmed overlap separately from unresolved possible overlap. If no reliable cross-cohort identifier or recruitment record exists, state that limitation rather than interpreting “no match found” as “no overlap.”
Specify that the target is a distinct person, and record separately the numbers of participants, specimens, visits, or assays where relevant.
Resolve repeated records and document how multiple specimens or visits map to one participant before calculating cross-cohort counts.
Use appropriately authorized common identifiers or a protected linkage process, and retain only the information needed for the overlap assessment.
Separate confirmed matches, ruled-out candidates, and unresolved cases; give the denominator and the method used to calculate overlap.
The overlap proportion needs a stated denominator. The fraction \(k/|V|\) answers what share of validation participants also appear in discovery. If the question instead concerns the share of all distinct people represented in both cohorts, the Jaccard overlap is \(k/|D\cup V|\). These summaries are not interchangeable. In the preceding example, that second quantity is \(6/224=0.0268\), or \(2.68\%\), which is smaller because its denominator includes people in both cohorts who are not shared.
Worked Example: Handling Candidate and Unresolved Matches
In a hypothetical validation cohort of 60 distinct participants, three people have exact matches to discovery records. A linkage procedure also flags four candidate pairs. A permitted review of recruitment records confirms two of those pairs, rules out one, and cannot resolve the remaining pair. Assume the exact matches and confirmed pairs are distinct validation participants, and that no other matches were identified.
The confirmed overlap is \(3+2=5\) participants. Its proportion is \[ \frac{5}{60}=0.0833, \] or \(8.33\%\). One additional participant remains a possible, unresolved match. If that candidate is a true match, the overlap is six; if not, it is five. The corresponding range is \[ \frac{5}{60}=8.33\% \quad \text{to} \quad \frac{6}{60}=10.0\%. \]
The defensible report is therefore five confirmed matches, one unresolved possible match, and a confirmed-overlap proportion of \(8.33\%\), with a possible-overlap range up to \(10.0\%\) under the stated candidate assessment. It would be misleading either to report six confirmed matches or to omit the unresolved candidate. The result also depends on the linkage procedure's ability to find true matches, so the absence of further candidates should not be described as a guarantee of zero additional overlap.
Why Overlap Changes the Validation Claim
If a person contributes to both discovery and validation, some information about that person may already have influenced the discoveries being evaluated. The validation result is then not wholly independent of discovery, even if the validation outcome was measured later or on a different specimen. This can weaken the interpretation of validation as evidence that a finding generalizes to new participants. The size and direction of any effect on a particular estimate or test depend on the analysis and data-generating process; an overlap count alone does not quantify the resulting bias.
The best handling depends on the validation objective and the analysis. If feasible and appropriate, a confirmed overlapping participant can be excluded from the validation analysis, with the resulting sample size and exclusion rule reported. In other settings, exclusion may change the target population or be inappropriate; the overlap should then be disclosed and its implications discussed. Removing one specimen while retaining another from the same person does not remove participant-level overlap. Likewise, splitting specimens from a participant between cohorts does not create independent participants.
An overlap assessment should also distinguish shared participants from shared physical specimens. A specimen-level duplicate may create direct reuse of the same measurement, while participant overlap can remain even when specimens are distinct. For multi-omics data, the same participant may appear in one cohort through genomic data and in another through transcriptomic or proteomic data. Checks should use the participant mapping across data layers, not just compare identical assay files.
Worked Example: Participants Are Not Specimens
Suppose a hypothetical discovery data set contains 48 biospecimen records from 40 distinct participants. A validation data set contains 35 biospecimen records from 30 distinct participants. Cross-cohort participant records confirm that four validation participants also occur in discovery; two of those four contributed different specimens in the two cohorts. Assume the participant mapping is reliable and the 30 validation participants have been deduplicated correctly.
The participant-level overlap count is four, and the validation overlap proportion is \[ \frac{4}{30}=0.1333, \] or \(13.33\%\). There are \(30-4=26\) validation participants with no confirmed discovery match. The fact that some overlapping participants contributed different specimens does not change the participant-level overlap count.
It would be incorrect to calculate \(4/35\) as the validation participant-overlap proportion, because 35 is the number of validation specimens, not participants. It would also be incorrect to claim that there are four duplicated specimens: the information given establishes four shared people, not four identical physical samples. The report should give both participant and specimen counts and state that the overlap identified is at the participant level.
Diagnostics, Comparisons, and Common Pitfalls
A useful diagnostic is to reconcile every count by unit. Check that the number of distinct participants is no greater than the number of participant-linked records, that confirmed shared participants are included in each cohort's participant total, and that the union count follows \(|D|+|V|-|D\cap V|\). Inspect candidate matches for systematic patterns, such as IDs that were reformatted in one cohort or many records sharing only a common site and birth year. Such patterns can reveal either missed matches or an overly broad linkage rule.
| Approach | Strength | Limitation |
|---|---|---|
| Exact comparison of common participant IDs | Simple and reproducible when IDs are consistently assigned and protected | Misses matches if IDs differ, are missing, or identify specimens rather than people |
| Recruitment-record review | Can resolve ambiguous identity and clarify enrollment history | Requires authorized access, careful documentation, and sometimes manual review |
| Probabilistic or rule-based record linkage | Can find candidate matches when direct IDs are unavailable or inconsistent | Can produce false matches and missed matches; results depend on fields and thresholds |
| Declaring no overlap without comparable records | No reliable advantage for establishing independence | Absence of detected matches is not evidence that overlap was adequately checked |
A common mistake is to reason, “The cohort names and study IDs differ, so the participants must be different.” Different labels do not establish different people: IDs may be study-specific, and a participant can enroll in both studies. The correction is to establish what each ID identifies, compare suitable records under the relevant permissions, and report any remaining uncertainty.
Other limitations should be stated plainly. Missing or inconsistent identifiers can prevent reliable linkage. A linkage process can miss participants whose records differ substantially, and aggressive matching can falsely join two people. Even a well-documented check applies only to the records and identifiers available; it cannot establish more than those data support. Report the matching unit, sources, rules, adjudication, confirmed count, unresolved count, and denominator. If overlapping participants are excluded, report the number excluded and the number remaining for validation.
Check Your Understanding
Answer each question, then compare your reasoning with the solutions.
- Discovery contains 100 distinct participants and validation contains 50. If four validation participants are confirmed in discovery, what is the validation overlap proportion and how many validation participants have no confirmed match?
- Why does a shared participant count not necessarily equal the number of duplicated specimens?
- A linkage procedure flags three possible matches, but review confirms one, rules out one, and leaves one unresolved. How should the confirmed overlap be reported?
- Is a validation cohort known to be independent if no identical study IDs are found? Explain.
Solutions. (1) The validation overlap proportion is \(4/50=0.08\), or \(8.00\%\). The number with no confirmed match is \(50-4=46\); this does not rule out unresolved or missed matches. (2) One person can contribute different specimens to the two cohorts, so participant identity does not imply the same physical specimen was reused. Conversely, a specimen record can be duplicated without representing a newly identified person. (3) Report one confirmed match, one ruled-out candidate, and one unresolved possible match. Do not count the unresolved case as confirmed; state that the overlap could be at least one and could be two under the reviewed candidates. (4) No. IDs may be cohort-specific, missing, or represent specimens rather than participants. Independence is not established by the absence of identical IDs; the report should describe what identifiers and records were checked and any limits on that check.