When Is Record Linkage Necessary in Research?

When Do You Need to Connect Records from Different Data Sources?

You have access to a dataset containing information relevant to your research question, but some important variables or events are stored elsewhere. A hospital database contains patient diagnoses but not deaths recorded after discharge. A university maintains enrollment records separately from graduate employment information. An administrative system captures service use but not outcomes recorded by another agency.

Should you link these datasets, or can the research proceed using the available records independently?

Record linkage may provide information that no individual dataset contains. However, connecting records introduces methodological challenges involving identity matching, incomplete coverage, privacy, and potential linkage errors.

The decision should therefore depend on whether linkage is needed to answer the research question and whether the resulting linked data can support valid conclusions.

Record Linkage Is Necessary When the Research Question Requires Information Distributed Across Records

Determining Whether Record Linkage Is Required and Feasible

What Is Record Linkage in Research?

Record linkage is the process of identifying and connecting records that refer to the same entity, such as a person, institution, household, organization, or event.

The records may originate from different databases or from multiple entries within the same database. Linkage can also be used to identify duplicate records, reconstruct longitudinal histories, or connect related events.

For example, a researcher may link hospital admission records with a mortality registry to determine whether patients died after discharge. Another researcher may connect university enrollment records with graduate survey responses to examine educational trajectories.

The objective is to establish which records belong together, not merely to place datasets in the same file.

Unlike a simple aggregation of institutional statistics, individual-level linkage attempts to preserve relationships between observations belonging to the same entity.

Distinguish Record Linkage from Ordinary Data Merging

Researchers sometimes use the terms merging, joining, and linking interchangeably. They overlap technically but can represent different methodological tasks.

Record linkage can include straightforward exact matching when reliable identifiers are available. More complex approaches are required when identity must be inferred from combinations of imperfect identifying information.

The distinction matters because uncertainty about whether two records belong to the same entity can introduce errors that ordinary database joins may not explicitly address.

When Is Record Linkage Methodologically Necessary?

Linkage becomes necessary when the intended research design requires information distributed across separate sources and the relationship between those sources is essential to the analysis.

Common situations include:

• Outcome ascertainment: linking baseline records to a separate source containing subsequent outcomes.
• Longitudinal follow-up: connecting observations collected at different times or by different organizations.
• Exposure measurement: identifying exposures recorded outside the primary outcome dataset.
• Confounding control: obtaining relevant baseline characteristics from complementary administrative sources.
• Population coverage: combining records to identify events or individuals omitted from a single system.
• Duplicate identification: determining whether multiple entries represent the same person or event.
• Cross-sector research: connecting educational, employment, healthcare, or other administrative records when the research question requires relationships between sectors.

These purposes have different implications for linkage quality. Missing a linked death record, for example, may affect outcome ascertainment, whereas incorrectly linking an exposure record may produce exposure misclassification.

Before requesting linkage, specify exactly which analytical variable or relationship cannot otherwise be established.

When Might Linkage Be Unnecessary?

Not every research question requires individual-level linkage.

If a single dataset already contains adequately measured exposures, outcomes, and necessary covariates, additional linkage may offer little methodological benefit.

Similarly, a study examining annual institutional enrollment totals may only require aggregate statistics. Connecting individual student records across systems would introduce unnecessary complexity if no individual-level relationships are being investigated.

Linkage may also be unnecessary when an existing, appropriately documented linked dataset already provides the required information.

The guiding principle is proportionality: obtain and connect only the information necessary to answer the research question.

Identify the Unit Being Linked

Before selecting a linkage method, establish what constitutes the same entity.

A person-level study requires identifying records belonging to the same individual. An event-level study may require determining whether two records describe the same hospitalization, transaction, or service encounter.

These tasks are not equivalent. A patient can have multiple admissions, and an admission can generate several records.

Researchers must also determine the expected relationship between datasets:

• One record in each source for the same entity.
• One individual linked to multiple events.
• Multiple records in both sources requiring entity resolution.
• Hierarchical relationships, such as patients within households or students within institutions.

Incorrect assumptions about these relationships can create duplicate observations, inflated event counts, or inappropriate analytical units.

Assess Whether Reliable Linkage Identifiers Are Available

Linkage requires information that can distinguish entities with sufficient reliability.

A unique identifier consistently assigned across databases may make matching relatively straightforward. However, identifiers can contain errors, be reused, change over time, or refer to different units across systems.

Where common identifiers are unavailable, combinations of identifying attributes may be used, subject to applicable permissions and safeguards.

Potential linkage fields include names, dates of birth, administrative identifiers, and other distinguishing characteristics. Their availability and permissible use vary considerably across settings and jurisdictions.

Researchers should evaluate the completeness, accuracy, uniqueness, and stability of identifiers rather than assume that matching fields are error-free.

For example, two people may share a name and date of birth, while the same person may appear under different names or inconsistent demographic details.

Deterministic vs. Probabilistic Record Linkage

Two broad approaches are commonly used to establish record correspondence.

Probabilistic linkage is often associated with the Fellegi-Sunter framework, which uses patterns of agreement and disagreement to help classify record pairs as matches, nonmatches, or potential matches requiring further assessment.

These approaches are not mutually exclusive. Practical linkage workflows may combine exact matching, probabilistic methods, and carefully controlled review procedures.

The appropriate method depends on identifier quality, dataset size, expected match relationships, available reference information, and the consequences of linkage errors.

Understand False Matches and Missed Matches

Record linkage is subject to two important types of error.

A false match can incorrectly assign an outcome, exposure, or other characteristic to the wrong individual. A missed match can cause an existing event to appear absent from the linked dataset.

For example, an incorrectly linked mortality record could classify a living participant as deceased, while a missed mortality linkage could make a deceased participant appear to have no recorded death.

The consequences depend on which variables are obtained through linkage and whether linkage errors differ across study groups.

Researchers should therefore evaluate linkage accuracy in relation to the intended analysis rather than regard the percentage of successfully linked records as a sufficient quality measure.

Why a High Linkage Rate Does Not Guarantee High Linkage Quality

The linkage rate usually describes the proportion of eligible source records linked to another dataset under a specified procedure.

A high linkage rate may indicate broad coverage, but it does not establish that the matches are correct.

For example, overly permissive matching rules may increase the proportion of linked records while also increasing false matches.

Conversely, conservative rules may produce highly credible accepted matches while missing some genuine relationships.

Linkage quality should therefore consider both match correctness and missed relationships, ideally using suitable validation information.

When a reference set of known matches and nonmatches is available, researchers may estimate sensitivity or recall, precision or positive predictive value, and other relevant performance measures.

The usefulness of these measures depends on how the reference set was constructed and whether it represents the records being linked.

Evaluate Whether Linkage Errors Could Be Systematic

Linkage errors are particularly concerning when they differ across population groups or institutions.

For example, records may be harder to link for individuals whose identifying information changes over time, contains transcription errors, or is recorded inconsistently across systems.

Differences in data completeness can produce unequal linkage success across age groups, service settings, or other relevant characteristics.

If linkage success is associated with both exposure and outcome, analyzing only successfully linked individuals may introduce selection bias.

Even when linkage is highly accurate overall, differential errors may distort subgroup comparisons.

These issues require more detailed consideration when evaluating how record linkage errors affect research design.

Determine Whether the Linked Sources Cover the Intended Population

Successful linkage does not establish complete population coverage.

Suppose a researcher links hospital records with an insurance database. Patients may receive services outside the insurance system, and some may have periods without observable coverage.

Even perfect matching between the two datasets cannot recover events absent from both.

Researchers should separately evaluate source coverage, linkage accuracy, and the completeness of the resulting analytical variables.

This distinction is especially important when linkage is intended to address missing care versus missing healthcare records.

Linkage can improve ascertainment, but it cannot transform incomplete source systems into a complete representation of reality.

Consider Privacy, Ethics, and Data Governance Before Linking

Record linkage may involve personally identifying information or combinations of attributes that create reidentification risks.

Researchers should establish an appropriate legal basis, institutional authorization, and applicable ethics review requirements before obtaining or processing linkage identifiers.

Requirements differ across jurisdictions, data types, institutions, and research purposes. Consent may be required in some circumstances, while other lawful research arrangements may permit linkage under specified safeguards.

Possible safeguards include:

• Separating identity linkage from substantive research analysis.
• Restricting access to direct identifiers.
• Using secure processing environments and approved data-transfer procedures.
• Applying data minimization and appropriate retention rules.
• Using pseudonymized analytical identifiers after matching.
• Documenting permitted uses and disclosure controls.

Pseudonymization does not necessarily make linked data anonymous. Combined datasets may remain identifiable through indirect characteristics or external information.

Privacy-preserving record linkage methods can reduce the need to disclose some identifying information to particular parties, but they introduce technical and governance considerations and do not eliminate every privacy risk.

Plan Linkage Before Finalizing the Research Design

Linkage feasibility should be evaluated early, preferably before the study's central analytical commitments are finalized.

Determine whether the necessary datasets exist, whether access can be authorized, which organization will perform matching, and whether suitable identifiers are available.

Researchers should also assess whether the required observation periods overlap and whether the linked sources use compatible definitions.

For example, linking a student enrollment database covering 2015–2025 with an employment database available only from 2023 may support some graduate outcome analyses but not equivalent follow-up for every graduation cohort.

Linkage also requires realistic consideration of governance approvals, secure infrastructure, technical expertise, and data preparation.

A research question that depends entirely on inaccessible linkage identifiers may be infeasible, even when both substantive datasets are otherwise available.

Document the Linkage Process for Reproducibility

Researchers should document the linkage units, source datasets, eligibility rules, identifier fields, matching procedures, accepted-match criteria, and treatment of uncertain or duplicate matches.

Report how many records were eligible, successfully linked, excluded, or left unmatched, using clearly defined denominators.

Where possible, describe validation procedures and assess differences between linked and unlinked records.

The RECORD statement and its extension for studies involving linked routinely collected data provide relevant guidance on reporting linkage methodology and quality.

Confidential identifiers and sensitive matching details need not be publicly disclosed when doing so would violate applicable safeguards. Methodological transparency should be achieved within legitimate privacy and security constraints.

Linking Hospital Admissions to Mortality Records

Suppose the linkage identifies 180 qualifying deaths that were not visible in the hospital database. Those additional outcomes demonstrate how linkage can materially improve ascertainment.

However, the observed number of linked deaths does not itself establish complete ascertainment. The final mortality estimate still depends on registry coverage, accurate matching, and the treatment of unresolved linkage uncertainty.

Common Mistakes When Deciding to Link Research Records

Decide Whether Linkage Is Necessary, Feasible, and Worth the Methodological Trade-Offs

Start with the research question rather than the availability of multiple datasets. Identify the information that cannot be obtained from the primary source and determine whether the intended analysis genuinely requires connecting records at the individual or event level.

Then assess the quality of the available identifiers, source coverage, expected linkage accuracy, and applicable data-governance requirements.

When linkage is required to address gaps in medical or administrative routine records, establish whether the additional source genuinely improves the intended measurement rather than merely adding another incomplete representation.

The strongest justification for linkage is not that it produces a larger dataset. It is that it makes an otherwise inadequately measured research question answerable under defensible assumptions.

What to Verify Before Beginning Record Linkage

Frequently Asked Questions About Record Linkage in Research

Link Records When the Research Question Requires Connections That One Dataset Cannot Provide

How to Cite This Guide

This guide is intended to be read, shared, and used in research, teaching, and academic work. If you draw on its ideas, explanations, or other content, please acknowledge the source by citing the guide. Doing so gives appropriate credit and helps your readers locate the original resource.