Subject Identifiers and Uniqueness

Why Subject Identity Needs an Explicit Rule

A subject identifier is the link that allows records from a study to be understood as belonging to the same participant. If that link changes between files, or if one identifier is accidentally assigned to two participants, later data preparation can combine different people or split one person into apparently separate subjects. Either error can affect data summaries and model inputs.

The identifier therefore needs a clear scope and a stable meaning. In this tutorial, a subject means the study participant as defined by the study’s data specification. An identifier is a label for that participant, not a measurement. The identifier may recur on many records for the same subject; subject-level uniqueness does not mean that the identifier must occur only once in a dataset.

The earlier material on Defining the Pharmacometric Record Grain establishes what one row represents. That distinction matters here: a subject identifier identifies the subject associated with a row, whereas a record-level identifier would distinguish individual rows. Repeated subject identifiers across multiple rows are usually expected. Do not test for subject-ID uniqueness across all observation rows as if every row had to represent a different participant.

Choose the Scope and Preserve the Identifier

Before checking data, specify the identifier’s scope. A study may assign IDs that are unique only within a site, within a study part, or within a particular source system. An ID such as “014” is not necessarily globally unique simply because it looks distinctive in one file. If different sites can issue the same local value, the site or another documented namespace is needed to interpret it.

A practical design uses a canonical study subject ID for analysis and retains the source identifiers and their provenance in a controlled mapping. The canonical value should be stable, distinct for each study subject, and assigned using a documented rule. It should not depend on row order or on the order in which files happen to be processed: sorting or adding a file could otherwise change assignments. Keep the raw source value as received so that transformations and reconciliations can be reviewed.

Identifiers are labels, even when they contain only digits. Converting “0007” to the number 7 discards information and may make it impossible to compare the value with source records. Likewise, changing case, removing punctuation, or trimming spaces can merge two distinct IDs if the source system treats those characters as meaningful. Apply only transformations authorized by the data specification, and record them. A blank string, a whitespace-only string, or a documented placeholder such as “UNKNOWN” may represent missingness; decide this by an explicit rule rather than by visual inspection alone.

Completeness and Uniqueness Checks

For a subject roster intended to contain one row per subject, let \(R\) be the set of roster rows and let \(I\) be the set of their interpreted subject identifiers. A complete, unique roster has one nonmissing identifier per row and a different identifier for every subject. Equivalently, if every row is expected to describe a distinct subject, then \(|I|=|R|\), provided no row has a missing identifier. This is a useful check, not proof that the identifiers refer to the correct people: a wrongly assigned but unique value can still pass.

For a file with repeated records per subject, the test is different. First check that every row that requires a subject association has a nonmissing subject identifier. Then compare the observed subject IDs with the authoritative roster or crosswalk. Repeated IDs across rows are allowed. A duplicate-ID test is meaningful only when the rows being tested are supposed to be one-per-subject, or when the test is scoped to a mapping table that should have one row per source identifier.

Missingness also requires a stated denominator. A rate calculated over data rows answers “what fraction of rows lacks a subject ID?” A rate over a subject roster answers “what fraction of expected subjects lacks an ID?” These are different questions. Report which one is being calculated, and distinguish genuinely missing values from values that merely look unusual.

Reconcile IDs Across Files

A repeated identifier in two files does not automatically indicate an error. If both files contain records for the same participant, the repetition is expected. The question is whether each source value has the same meaning in each file and whether the relationship to the canonical study subject is consistent. A useful crosswalk records the source system or file context, the raw ID, the canonical subject ID, and the evidence or rule supporting the mapping.

For a given source namespace, one source ID should not map to two canonical subjects. Conversely, the same canonical subject may have different IDs in different systems, such as a laboratory system and an enrollment system. That is not a collision if the mappings have been verified. Do not infer identity from similar demographics alone; use the approved linkage evidence and refer unresolved discrepancies to the responsible data owner.

Why the Roster Check Works—and What It Cannot Prove

The roster check follows directly from the definition of a one-row-per-subject roster. Let \(f\) assign to each roster row its usable identifier. If \(f\) is injective, then distinct rows have distinct identifiers. If identifiers are also present on every row, the number of distinct identifiers equals the number of rows. Conversely, if two rows share an identifier, \(f\) is not injective and the roster cannot distinguish those rows using that identifier.

The condition is necessary for a correctly constructed roster, but it is not sufficient to verify identity. For example, swapping two participants’ otherwise unique IDs preserves the number of distinct values while assigning IDs incorrectly. Automated checks establish structural properties of the values; comparison with authoritative source records establishes whether the assignments are correct.

Diagnostics, Alternatives, and Limitations

Useful diagnostics include counts of blank and placeholder values, counts of distinct IDs within each stated scope, duplicate rows in a one-row-per-subject roster, and source IDs with more than one canonical mapping. Review both exceptions and the expected overlap between files. A clean count can still hide a wrong mapping, so retain an audit trail of corrections and the source evidence used to make them.

There are several reasonable identifier designs, but each has limits:

• Use a source-local ID alone. This is simple when the source guarantees uniqueness over the full intended scope. It fails when IDs can be reused by sites, systems, or study parts.
• Combine the local ID with a namespace. A site or source-system prefix can distinguish reused local values if the namespace is reliable and stable. Inconsistent site labels or missing context can still create ambiguity.
• Assign a canonical study ID and keep a crosswalk. This supports harmonization across sources and keeps source values available for tracing. The crosswalk itself must be access-controlled as appropriate, complete, and versioned; a canonical ID does not correct an unsupported source linkage.

Common failure modes include converting IDs to numbers, treating every repeated subject ID in a record file as a duplicate error, assuming the same raw ID always means the same subject across sites, and silently assigning a new ID to a row with a missing value. Each can either erase distinguishing information or create a false link. A further limitation is that a uniqueness test cannot detect an ID that is unique but belongs to the wrong participant. That requires reconciliation with authoritative records.