Describing Data Sources and Record Linkage

Why Data Provenance Matters

A comparative analysis can be calculated correctly and still be difficult to assess if readers cannot tell where its records came from or how those records were assembled. A source name alone is not enough: a hospital record system, insurance claims database, and disease registry may cover different people, periods, and events, even when they contain similarly named variables. When information from multiple sources is linked, readers also need to know how the linkage was performed and what evidence supports its accuracy.

For routinely collected health data, the RECORD guidance extends STROBE reporting recommendations with attention to data sources, access, coding, linkage, and validation. The aim is not to provide an inventory of every technical detail. It is to report enough for a reader to understand the data’s origins and limitations, judge the relevance of the sources to the study question, and assess whether the linkage process could have introduced error.

Describe Each Source, Not Just Its Name

For each source, report its name and the organization responsible for maintaining it, where that information can be shared. Identify the type of source—for example, administrative claims, an electronic health record system, a laboratory information system, or a registry—and explain its primary collection purpose. A source designed for billing, care delivery, or surveillance may record the same clinical event differently because it was created for a different operational purpose.

Describe the source’s relevant coverage: the geography, participating facilities or organizations, population or service setting represented, and calendar period covered. Distinguish the database’s overall coverage from the records available to the analysis. Report material gaps, such as nonparticipating facilities, changes in participation over time, or populations whose care may occur outside the source. These are descriptions of what the source can capture, not conclusions about who qualifies for the analysis.

Explain the provenance of the analysis data as well as the original source. State how the data were obtained or accessed, such as through a data custodian or a secure research environment, and describe major processing steps relevant to interpreting the records. If a source is updated, say which version or extraction period was used when available. Report the source’s collection purpose even if it seems obvious: a claims database’s records, for example, are generated in a reimbursement context, not necessarily to create a complete clinical history.

Coding Systems and Operational Definitions

Routine data often represent diagnoses, procedures, medications, or other events using codes. Report the coding system and version where relevant, the time period for which it applied, and how the codes were selected. If an analysis uses a published code list, cite or otherwise identify it. If investigators created or modified a list, describe the process and provide the list or an accessible supplement when permitted. Explain whether codes were used to identify a recorded event, a service, or a classification; the presence of a code does not automatically establish that a clinical condition was confirmed.

When a definition depends on an algorithm rather than a single code, describe the parts that affect interpretation—for example, whether one or multiple records were required, whether code position mattered, and how the algorithm treated conflicting records. Do not substitute a broad statement such as “standard diagnostic codes were used” for the system, version, and operational choices a reader would need to understand the definition. This reporting is about how source data were represented and interpreted, not a substitute for describing the study’s later analytical decisions.

Explain How Records Were Linked

For linked data, describe which sources were linked, the unit being matched, the identifiers or identifier classes used, and whether linkage was deterministic, probabilistic, or a combination. Deterministic linkage applies specified rules to identifiers, such as exact agreement on a unique identifier. Probabilistic linkage combines evidence across identifiers that may be incomplete or inconsistent, assigning scores or probabilities to candidate pairs and applying a decision rule. State the main rules or decision thresholds and how ambiguous or conflicting candidates were handled.

Report who performed the linkage and where, to the extent possible without compromising privacy or security. Explain whether identifiers were encrypted, transformed, or handled by a trusted third party, and what information was returned to the analysts. Privacy-preserving linkage can reduce exposure of direct identifiers, but it does not by itself establish that the links are accurate. Also state the denominator used for any linkage rate: for example, the number of records submitted for linkage, rather than an unspecified “study population.”

Validation Evidence: What Was Checked?

Validation asks whether source definitions or linkage decisions correspond adequately to a more credible reference for the intended purpose. Report what was checked, how the reference was established, who performed the review if relevant, and the results with clear denominators. For linkage, reference labels might come from manual review of records using information unavailable to the linkage algorithm. For coded definitions, the reference might be a clinical review or another appropriate source. A validation result applies to the definition and population that were checked; it is not automatically evidence of accuracy in every setting or period.

For binary linkage decisions, a validation sample with a reference classification can be summarized using four counts: true matches accepted by the procedure, nonmatches accepted, true matches rejected, and nonmatches rejected. Sensitivity is the proportion of reference true matches accepted; positive predictive value is the proportion of accepted links that are true matches. Specificity is the proportion of reference nonmatches rejected. These measures answer different questions, so “linkage accuracy” without a measure and denominator is ambiguous.

A Practical Reporting Sequence

A compact, reproducible account can be built by working through the following sequence. The sequence separates the origin and properties of each source from the linkage and the evidence used to assess it.

Diagnostics, Alternatives, and Limitations

Several warning signs indicate that a source or linkage description is too thin: a source is identified only by a familiar acronym; coverage is described as “national” without saying which people or organizations contribute; codes are called “standard” without a system or version; or a linkage rate is reported without its denominator. Another warning sign is reporting only a high linkage rate as evidence of correct linkage. A procedure can link many records and still accept false matches, or achieve a low false-match rate while missing many true matches.

The appropriate linkage approach depends on the identifiers available and the risks of errors. Exact deterministic linkage may be straightforward when a reliable unique identifier is shared, but exact agreement rules can miss genuine matches when identifiers are missing or inconsistent. Probabilistic linkage can use partial agreement across fields, but depends on specified scoring and decision rules and may leave uncertain pairs unresolved. A hybrid may apply deterministic rules to clear cases and probabilistic rules to the remainder. These are alternatives in method, not guarantees of quality: report the method actually used and evidence relevant to its performance.

Single-source analyses avoid cross-source matching errors, but a single database can still have incomplete coverage, measurement limitations, or coding errors. Linked data can add information and broaden observed histories, but can also introduce false matches, missed matches, and selection into the linked subset. RECORD-style reporting makes these issues visible; it cannot repair them. Validation may be infeasible, unavailable, or limited to a small sample, and results may not transport across populations, time periods, or coding changes. State those limits rather than implying certainty.

A useful justification for reporting both linkage yield and validation evidence follows directly from their definitions. The linkage rate uses all records submitted as its denominator, whereas positive predictive value uses only accepted links. Since these denominators differ, a high value for the first ratio does not mathematically determine the second. The two quantities answer separate questions—how often the procedure returns a link, and how often its accepted links are correct—and should not be used interchangeably.