Harmonizing Phenotype Definitions Across Recruitment Sites

Why Phenotype Harmonization Matters

A multi-site genome-wide association study can collect data from thousands of people yet still fail to measure the same phenotype at every site. One site may define a case from a specialist diagnosis, another from a billing code, and a third from a questionnaire. If the same label is attached to all three, it does not make their meanings equivalent. A shared phenotype definition must specify what evidence counts, when it counts, and how conflicting or incomplete records are handled.

Harmonization is a design task: it converts site-specific records into a common, explicit phenotype specification before the data are analyzed. The goal is not to erase meaningful differences among sites or to force incomparable records into a single category. It is to make the intended construct and the rules for representing it clear, consistent, and auditable.

Definition: A harmonized phenotype is a phenotype represented across sites by a common operational specification: the target construct, eligible sources of evidence, diagnostic or measurement rules, assessment window, and rules for assigning cases, controls, and indeterminate records. The raw data and collection procedures may differ by site, but each site’s records must be mapped to the shared specification in a documented way.

What a Common Specification Must Decide

Start by naming the clinical or biological construct and the study’s intended estimand. The earlier material on choosing GWAS phenotypes and estimands distinguishes the construct of interest from the particular records used to represent it. Harmonization addresses that second step: how each site’s data will operationalize the construct. A useful specification makes at least four decisions explicit.

Diagnostic evidence: Which code systems, code versions, clinician diagnoses, medication records, laboratory criteria, or participant reports can support a case classification? Are all sources equivalent, or does one take precedence?
Measurement protocol: For a quantitative or repeatedly measured trait, which instrument, specimen, units, preparation conditions, replicate summary, and quality checks define an acceptable value?
Assessment window: What event anchors assessment, such as enrollment or a clinical visit, and which dates or ages are in scope? Specify how to choose among multiple eligible observations.
Discordance and missingness: How are conflicting records resolved, and what happens when evidence is absent, incomplete, or outside the window? Absence of a recorded diagnosis is not automatically evidence that a person is a control.

The specification should also identify the source and version of each site’s data, the mapping from local values to common categories, and any site where the required evidence cannot be obtained. Keep three outcomes available when appropriate: case, control, and indeterminate. The indeterminate category prevents “not classifiable from available records” from being silently treated as “unaffected.”

Worked Examples of a Shared Operational Definition

Worked Example: Mapping Local Diagnostic Records

Suppose two hypothetical recruitment sites contribute records for a study of a chronic inflammatory condition. The study team specifies that a case requires either a confirmed specialist diagnosis in the record or at least two qualifying diagnosis entries on separate dates. A control requires an adequate review period with no qualifying evidence and no relevant specialist diagnosis. All other records are indeterminate. The code lists below are hypothetical mappings created for this example, not clinical coding guidance.

Site A uses code system A and Site B uses code system B. A crosswalk maps A17 and B04 to “qualifying diagnosis entry,” and maps A90 and B91 to “specialist-confirmed diagnosis.” The crosswalk is reviewed against the code descriptions and versions used at each site. A code not on the approved list is not assumed to be equivalent just because its label sounds similar.

At Site A, participant 1 has A17 entries on two distinct dates and no specialist record. The two entries meet the stated rule, so this participant is a case. Participant 2 has one A17 entry and a documented review period of four years with no further qualifying evidence. If the protocol requires at least three years of review for a control, participant 2 is a control. Participant 3 has one A17 entry but only six months of available follow-up; there is not enough evidence for either the case rule or the control rule, so the participant is indeterminate.

At Site B, participant 4 has one B91 specialist-confirmed diagnosis. This satisfies the first case rule and the participant is a case, even without a second code entry. Participant 5 has no qualifying code, but the record does not establish the required review period. That participant is indeterminate, not a control.

This mapping retains different local code systems while applying the same decision rule. The review-period requirement matters: without it, a site with shorter or less complete records could produce many apparent controls simply because it had fewer opportunities to observe a diagnosis.

Worked Example: Aligning a Quantitative Measurement

Consider a hypothetical study that defines its phenotype as the nearest eligible LDL cholesterol measurement within 90 days of enrollment, recorded in mg/dL, with the measurement date and fasting status retained. Site A reports mg/dL; Site B reports mmol/L. The protocol specifies the conversion \(1\ \mathrm{mmol/L}=38.67\ \mathrm{mg/dL}\). This conversion aligns units, but it does not make different specimen conditions or instruments automatically interchangeable.

A Site A participant has an eligible result of 124 mg/dL on day 18. A Site B participant has an eligible result of 3.0 mmol/L on day 12. The converted value is:

3.0\ \mathrm{mmol/L}\times 38.67\ \frac{\mathrm{mg/dL}}{\mathrm{mmol/L}} =116.01\ \mathrm{mg/dL}.

Rounded to three significant figures, the Site B value is 116 mg/dL. Both records fall within the 90-day window, so each supplies an eligible value under the stated rule. The analysis dataset should retain the source unit, converted value, date relative to enrollment, and fasting status; this makes the harmonization auditable and allows protocol differences to be assessed.

Now suppose Site B’s result was obtained on day 105. It falls outside the specified window, so it cannot serve as the primary harmonized measurement. The participant is missing that phenotype under this definition unless another eligible result exists. The team should not extend the window for that participant after seeing the value. If fasting status or assay method differs materially between sites, the protocol must decide whether those measurements remain eligible, require calibration or a separate indicator, or cannot be combined as the same phenotype. A unit conversion alone does not resolve such differences.

Worked Example: Resolving Conflicting Records

A hypothetical study defines a case as a specialist-confirmed diagnosis recorded by the end of the 90-day post-enrollment window. A control must have a complete review through that window, no qualifying diagnosis, and no disease-specific treatment record. A medication record alone is not sufficient to establish a case because the medication can have other indications. The protocol says a specialist diagnosis takes precedence over a questionnaire response.

Participant A has a specialist diagnosis on day 40 and answers “No” to the enrollment questionnaire. The specified precedence rule makes this participant a case: the structured clinical evidence meets the case definition, while the questionnaire response does not override it.

Participant B has no diagnosis code, answers “Yes” to the questionnaire, and has a medication record that could have more than one indication. No source satisfies the case rule, and the participant’s report conflicts with the other available evidence. Under this specification, Participant B is indeterminate pending adjudication; the medication alone does not resolve the conflict.

Participant C has no qualifying diagnosis or treatment record and answers “No,” but records end on day 30. Since the required review through day 90 is incomplete, C is indeterminate rather than a control. These classifications follow the prespecified rules; they do not depend on genotype or on whether a participant’s classification improves an association result.

Why Site-Specific Differences Can Change the Result

A simple calculation shows why common rules matter even when the underlying phenotype has no genetic association. Suppose two sites have different probabilities of recording a case under their local definitions: 0.10 at Site A and 0.30 at Site B. Assume, for illustration, that the true phenotype distribution is the same across genetic groups and the difference arises only from site-specific classification. If 80% of the \(G=0\) group but only 20% of the \(G=1\) group comes from Site A, the expected recorded case proportions are:

P(Y^*=1\mid G=0)=0.80(0.10)+0.20(0.30)=0.14,

P(Y^*=1\mid G=1)=0.20(0.10)+0.80(0.30)=0.26.

The recorded proportions differ by \(0.26-0.14=0.12\), despite the assumed absence of a true difference between genetic groups. Here \(Y^*\) denotes the recorded classification. The calculation is not a general model for every GWAS; it isolates one design problem: when sites classify differently and their contributions differ across groups, inconsistent definitions can create or distort an apparent association. A shared specification reduces this avoidable source of variation, though it does not by itself remove other sources of confounding or measurement error.

This is why harmonization is more than a data-cleaning step. It determines which participants are counted as cases and controls, and therefore changes the phenotype to which an association result refers. The site-level implementation should be checked before association analysis, not adjusted until results appear to agree.

A Practical Harmonization Workflow

Write the common target definition.
Name the construct, target population context, assessment anchor, and intended case, control, or quantitative-trait representation. State what does not qualify.

Inventory site-level evidence.
For every site, document source systems, code versions, available dates, instruments, units, collection conditions, and missing-data patterns. Do not infer equivalence from similar labels.

Build and review the mapping.
Map local codes or measurements to the common specification. Record unsupported or ambiguous values, conversion rules, and protocol differences that may affect comparability.

Prespecify resolution and eligibility rules.
Define windows, precedence among sources, duplicate handling, adjudication, and indeterminate status. Apply the same logic at every site wherever the available evidence permits.

Validate classifications and document limitations.
Review representative records, compare category counts by site, and inspect reasons for indeterminate or excluded records. Record deviations and sites where the common definition could not be implemented.

Validation should check the implementation, not optimize the phenotype based on genetic results. For example, a sample of records can be reviewed to confirm that code mappings and date rules were applied as written. Site counts and missingness summaries can reveal an unexpected mapping error or a site that lacks the required follow-up. If the shared definition cannot be supported at one site, state that limitation and make an explicit analysis-design decision rather than silently substituting a different definition.

Common Pitfalls

Assuming one label means one phenotype. A shared variable name does not establish that sites used equivalent diagnostic evidence or measurement procedures. Compare the operational rules, not just the labels.
Treating no record as a negative record. A participant without a diagnosis entry may have incomplete follow-up or an unreviewed source. Classify as a control only when the specified control evidence is available.
Forcing discordant records into case or control. If the evidence does not meet either rule, use an indeterminate category or a prespecified adjudication process. Do not resolve ambiguity by choosing whichever classification is convenient.
Assuming unit conversion solves protocol differences. Converting units corrects scale, not differences in timing, specimen conditions, instruments, or measurement procedures.
Changing the definition after examining associations. Data-driven changes made to improve or weaken genetic results undermine interpretability. Define and document the mapping before association results are used to assess it.

Key takeaway: A defensible cross-site phenotype is a common set of explicit rules, not merely a common name. Align the evidence, measurement conditions, assessment window, and discordance handling; retain indeterminate status when the records do not support a reliable case or control classification.

Check Your Understanding

Use the stated rules and distinguish a true negative classification from insufficient evidence.

Why is a shared phenotype label insufficient to establish that two recruitment sites measured the same construct?
A participant has no qualifying diagnosis code, but the available record covers only two months when the control rule requires two years. Should the participant be classified as a control? Explain.
A quantitative measurement is converted to the common unit but was collected outside the prespecified assessment window. Is it eligible under the rule in the measurement example?
What should a harmonization specification say about a medication record that can have several indications?
In the site-mixture calculation, what design feature produced different recorded case proportions across genetic groups despite the assumed absence of a true phenotype difference?