How Do Record Linkage Errors Affect Research Design?

What Happens to Your Study When Records Are Linked Incorrectly?

You have connected patient records to a mortality registry, student records to employment outcomes, or administrative records across institutions. The linkage process reports a high matching rate, and the combined dataset appears ready for analysis.

But some records may have been linked to the wrong individuals, while other records belonging to the same person may not have been linked at all. These errors can affect which participants enter the study, how exposures and outcomes are classified, and whether the estimated associations are trustworthy.

Should you exclude uncertain matches, adjust the statistical model, or redesign the study? The answer depends on the type of linkage error, the variables obtained through linkage, and how those errors relate to the research question.

Linkage Errors Can Introduce Misclassification and Selection Bias

How Linkage Errors Influence Research Validity

Understand the Two Fundamental Types of Linkage Error

Record linkage attempts to establish which observations refer to the same person, organization, event, or other entity. Errors occur when this correspondence is determined incorrectly.

These errors can have different consequences. A false match may assign someone else's outcome to a participant. A missed match may leave an existing outcome unobserved or exclude an eligible participant from the analytical dataset.

For example, when hospital records are linked to mortality records, a false match could assign a death to the wrong patient. A missed match could leave a genuine death unrecognized.

Neither error is necessarily visible in the final analytical dataset. Once a match is accepted, its information may appear as an ordinary observation unless linkage uncertainty has been preserved.

Why Linkage Rate and Linkage Accuracy Are Not the Same

A linkage rate describes the proportion of eligible records successfully linked under a specified procedure. It does not establish whether the accepted links are correct.

A permissive matching procedure may produce a high linkage rate while accepting more false matches. A restrictive procedure may produce fewer false matches but fail to identify some genuine relationships.

When suitable reference information is available, researchers can evaluate linkage using measures such as:

• Precision or positive predictive value: the proportion of accepted matches that are genuine.
• Recall or sensitivity: the proportion of genuine matching relationships successfully identified.
• False-match rate: a measure of incorrectly accepted links, with the denominator explicitly defined.
• Missed-match rate: the proportion of genuine matching relationships not identified, when those relationships can be established.

Terminology and denominators vary across linkage evaluations. For example, the proportion of accepted matches that are false is not necessarily the same as the false-positive rate among all nonmatching candidate pairs.

Report the exact measure and denominator rather than relying on an ambiguous statement that linkage was highly accurate.

How False Matches Affect Exposure and Outcome Measurement

False matches can transfer information from one individual to another.

If exposure information comes from one dataset and outcomes from another, an incorrect link can create an exposure-outcome combination that never existed for either person.

For example, a student receiving a scholarship could be incorrectly linked to another graduate's employment record. The resulting dataset would contain a false relationship between scholarship receipt and employment status.

In healthcare research, an incorrectly linked diagnosis, medication record, or death certificate may misclassify a central analytical variable.

The consequences depend on the distribution of the linked variables and how matching errors occur. Misclassification may attenuate an association, exaggerate it, or change its direction.

Researchers should not assume that false matches necessarily bias estimates toward the null.

How Missed Matches Affect Outcome Ascertainment

Missed matches can make existing events appear absent.

Suppose a hospital cohort is linked to a mortality registry. If a genuine death record is not linked to the correct patient, the analytical dataset may incorrectly classify that patient as having no recorded death.

The problem becomes especially consequential when the research design treats unmatched outcome records as negative outcomes.

Before using this rule, researchers should establish whether the target database adequately captures the relevant events and whether linkage sensitivity is sufficient for the intended inference.

This resembles the broader problem of treating the absence of a recorded diagnosis as evidence of disease absence. In both situations, a missing record may reflect incomplete ascertainment rather than a genuinely negative condition.

How Linkage Errors Can Create Selection Bias

Selection bias can arise when inclusion in the analytical sample depends on successful linkage and linkage success is related to characteristics relevant to the research question.

For example, individuals with incomplete identifiers, inconsistent names, or fragmented administrative histories may be less likely to link successfully.

If these characteristics are also associated with an exposure or outcome, restricting analysis to successfully linked records may produce a systematically selected sample.

Consider a study of educational attainment and employment. If graduates with unstable administrative records are less likely to be linked to employment data, excluding them may alter the estimated employment outcomes of the graduate population.

Even accurate matches among retained participants do not eliminate bias introduced by excluding people who could not be linked.

Researchers should therefore compare linked and unlinked records using characteristics available before linkage, while recognizing that such comparisons cannot reveal every unobserved difference.

Distinguish Differential from Nondifferential Linkage Error

Linkage errors may occur with different probabilities across exposure groups, outcomes, institutions, or demographic characteristics.

Differential error may arise when one institution maintains more complete identifiers than another, or when certain population groups experience more frequent changes in identifying information.

For example, if patients receiving a particular treatment have more complete records, their outcomes may be linked more successfully than outcomes among untreated patients.

This can create apparent differences in outcome occurrence even when the underlying event rates are similar.

Nondifferential error is not automatically harmless. Its effects depend on the linkage process, the affected variables, and the statistical estimand.

Consider the Role of Linkage Errors in Different Research Designs

The same linkage procedure may therefore be adequate for one research purpose but insufficient for another.

A small number of false links may be especially consequential when the outcome is rare, while missed links may substantially affect studies requiring complete longitudinal follow-up.

Evaluate Linkage Quality Before Finalizing the Analytical Sample

Researchers should examine the linkage process before treating the linked dataset as an ordinary complete sample.

Begin by documenting the source populations, matching identifiers, eligibility rules, linkage procedures, and handling of uncertain matches.

Where feasible, obtain a validation dataset containing independently established matching relationships. Such information can support estimates of false and missed matches.

Validation may involve clerical review, trusted identifiers, or independent reference information. However, clerical review is not necessarily an error-free reference standard, and validation samples may be unrepresentative.

Review linkage performance across relevant institutions, periods, and population groups.

If linkage quality cannot be adequately evaluated, this uncertainty should influence both study design and the strength of the conclusions.

Should Uncertain Matches Be Accepted, Rejected, or Reviewed?

Matching algorithms often produce records whose correspondence is uncertain.

Accepting more uncertain links may reduce missed matches but increase false matches. Rejecting them may improve confidence in accepted links while excluding genuine relationships.

Some linkage workflows use an intermediate category for possible matches requiring additional review.

The appropriate decision depends on available identifiers, validation evidence, privacy constraints, and the relative consequences of the two error types.

For example, false attribution of a rare adverse outcome may have different analytical consequences from failing to connect a routine administrative transaction.

Matching thresholds should be selected using defensible criteria rather than adjusted solely to produce a preferred sample size or substantive result.

Can Statistical Methods Correct Linkage Errors?

Some methods can account for linkage uncertainty when suitable information about matching probabilities or error processes is available.

Depending on the setting, researchers may consider probabilistic analyses, multiple plausible linkage configurations, weighting approaches, or statistical models that incorporate uncertainty about record correspondence.

These approaches require assumptions about the linkage process and sufficient information to estimate relevant error characteristics.

They are not universal corrections. A regression model that includes demographic covariates does not necessarily correct outcomes assigned to the wrong individuals.

Similarly, treating all accepted links as certain can underestimate uncertainty when linkage errors are consequential.

Researchers should distinguish conventional sampling uncertainty from uncertainty introduced by the linkage process.

Use Sensitivity Analyses to Examine the Consequences of Linkage Decisions

Sensitivity analyses can help determine whether substantive conclusions depend on matching rules or assumptions about linkage errors.

Possible analyses include comparing results under more restrictive and more permissive matching criteria, evaluating alternative treatment of uncertain links, or modeling plausible levels of false and missed matches.

When relevant, researchers may examine whether conclusions change after restricting analysis to records with high-confidence links.

However, restricting to high-confidence matches may introduce selection bias if match confidence is related to population characteristics.

Agreement across alternative linkage strategies can provide evidence of robustness to the assumptions examined, but it does not prove that the linked data are error-free.

When Should Linkage Problems Change the Research Design?

Linkage quality should influence design decisions before the main analysis begins.

If essential outcomes cannot be linked with sufficient reliability, a cohort study may require a different outcome source, narrower population, or revised follow-up strategy.

If linkage success differs substantially across institutions, direct comparisons may be difficult to justify without additional validation or adjustment.

If the available identifiers cannot distinguish individuals reliably, individual-level linkage may not be feasible.

In such circumstances, reconsidering the research design may be more defensible than applying increasingly elaborate statistical methods to poorly established relationships.

The initial decision about whether record linkage is necessary should therefore include an assessment of whether the intended linkage can be performed with adequate quality.

Report Linkage Quality and Its Implications Transparently

Report the numbers of eligible records, accepted matches, unmatched records, duplicates, and uncertain matches, using explicit denominators.

Describe the linkage method, relevant identifiers, matching rules, validation procedures, and available quality measures, subject to legitimate privacy and security restrictions.

Explain whether linkage errors could affect the primary exposure, outcome, population selection, or subgroup comparisons.

Where appropriate, report sensitivity analyses and distinguish observed findings from assumptions about linkage accuracy.

The RECORD statement and RECORD-PE guidance provide relevant reporting principles for research using routinely collected and linked health data.

How Missed Mortality Links Can Distort a Cohort Study

In this hypothetical example, differential linkage sensitivity creates an apparent mortality difference where none exists in the reference data.

The lesson is not that every missed match produces this pattern. Rather, linkage errors can alter the direction and magnitude of estimated associations when their occurrence differs across study groups.

Common Misconceptions About Linkage Quality

Make Linkage Quality Part of the Research Design

Before finalizing the analytical dataset, identify which variables depend on linkage and how incorrect or missed links could affect the main conclusions.

Prioritize linkage-quality assessment for central outcomes, exposures, and eligibility decisions. Evaluate whether matching errors are likely to differ across relevant study groups.

Where linked records originate from institutions with different data-generation practices, evaluate institutional coding differences separately from linkage errors. Correctly matching records does not establish that the linked variables are measured equivalently.

The relevant question is whether the linked dataset supports the intended inference, not simply whether the matching procedure produced a sufficiently large sample.

What to Verify Before Analyzing Linked Records

Frequently Asked Questions About Record Linkage Errors

Linkage Accuracy Is a Research Validity Issue, Not Just a Technical Detail

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.