Why Patient-Level Linkage Matters
An external-control cohort may be assembled from several data systems, each of which stores records rather than a single, definitive row per person. One patient can appear in a hospital record, a claims record, and a registry; a patient can also appear more than once within a source. If those rows are treated as different people, the constructed cohort no longer represents the intended patient-level population. If records from different people are mistakenly combined, distinct patients disappear from it.
Linkage is therefore part of defining the analysis population, not just a data-cleaning convenience. The relevant unit for a patient-level comparison is usually the unique eligible patient, while the source data may contain multiple records or enrollment episodes for that patient. The analyst must specify what constitutes a patient, which records belong to that patient, and whether repeated entries are duplicate representations or distinct episodes that need to be retained. These choices affect the number of patients and the number of patients with an outcome.
Before linking, state the unit the analysis requires. For example, a patient-level risk analysis might require one eligible patient per person, with an outcome indicator that records whether that patient experienced the specified event. A study of recurrent events, by contrast, may need to retain multiple events for one person. Removing every repeated row without distinguishing patients, enrollment episodes, and events can destroy information the analysis needs.
Linkage Decisions and Their Assumptions
A linkage procedure compares identifying fields such as date of birth, name, address, or a source-specific identifier. Exact agreement on a reliable, unique identifier can support deterministic linkage. When identifiers are incomplete, misspelled, outdated, or recorded differently across systems, probabilistic linkage combines the evidence across fields. Either approach requires assumptions about identifier quality, how fields are recorded, and whether candidate records can be compared reliably.
For a candidate pair of records, let \(M\) denote the event that they refer to the same person, and let \(A\) denote the observed pattern of agreement and disagreement across the fields being compared. A probabilistic method can summarize the evidence using the likelihood ratio: how much more likely the observed pattern is for a true match than for a nonmatch. Bayes’ rule combines that evidence with the prior probability that a candidate pair is a match.
The second factor is the likelihood ratio. A rare agreement, such as an exact match on a reliable identifier, can provide stronger evidence than agreement on a common value such as sex. A disagreement can count against a match, but its weight depends on the field: a current address may change, while a valid date of birth usually should not. If the method multiplies field-specific likelihood ratios, it assumes the fields are conditionally independent given match status, or uses a model that otherwise accounts for their dependence. Treating correlated fields as independent can count the same evidence more than once and produce overconfident match probabilities.
For a decision-based linkage, a threshold may be used: pairs with match probability above the threshold are linked, pairs below it are not, and intermediate cases are reviewed or left unresolved. A threshold is a decision rule, not a fact about identity. The appropriate rule depends on the relative consequences of falsely merging two patients and failing to combine records from the same patient, as well as on any constraints such as one-to-one linkage between sources.
Worked Example: Updating the Probability of a Match
Suppose a record pair has a prior match probability of 0.01. In a hypothetical linkage model, the observed date-of-birth agreement has a likelihood ratio of 20, and sex agreement has a likelihood ratio of 1.1. Assume for this calculation that the two fields are conditionally independent given match status. The combined likelihood ratio is \(20\times1.1=22\).
The prior odds are \(0.01/0.99=0.01010\). The posterior odds are therefore \(0.01010\times22=0.2222\). Converting those odds to a probability gives:
Despite agreement on both fields, the posterior match probability is only about 18.2%. The low prior probability matters: most candidate pairs were nonmatches before examining these fields. This pair should not be declared a match solely because the fields agree. The calculation is only as credible as its prior probability, likelihood ratios, and conditional-independence assumption; a production procedure would also need to consider the candidate-generation process and any linkage constraints.
From Linked Records to a Patient-Level Cohort
A defensible cohort-building process preserves the source records and the linkage decisions rather than overwriting them. It should document which identifiers were compared, which pairs were accepted, rejected, or unresolved, the rule or model used, and how each linked set of records became an analysis unit. This makes it possible to trace a patient-level row back to its contributing records and to revise decisions when audits reveal errors.
State whether the cohort contains unique people, enrollment episodes, or another unit. Define how multiple episodes for one person are handled.
Standardize fields without erasing meaningful differences, identify unavailable or invalid values, and generate candidate record pairs using a documented rule.
Apply deterministic or probabilistic rules, review ambiguous pairs when feasible, and retain match probabilities or decision categories with the underlying records.
Assign records to patient identities, distinguish duplicate representations from distinct enrollments, and apply the prespecified cohort rules.
Compare record, patient, and episode counts; check event counts against source records; and investigate unexpected changes before analysis.
The distinction between duplicate representation and repeated enrollment is substantive. If a person has two enrollment rows but both describe the same eligible entry, counting both as patients inflates the cohort. If the rows represent separate episodes that the study intends to analyze, collapsing them without retaining the episode structure changes the target unit. The rule should be set from the study definition, not chosen after seeing which version produces a more favorable outcome result.
Worked Example: Quantifying the Effect of Duplicate Patients
In a hypothetical external-control cohort, 50 unique patients include 10 patients with the outcome, giving a patient-level outcome proportion of \(10/50=0.20\), or 20.0%. A linked extract instead contains 60 rows because 10 patients appear twice. Of those 10 duplicated patients, 4 had the outcome and 6 did not. Assume each duplicated row repeats the patient’s same outcome status.
If all rows are incorrectly counted as separate patients, the row-based proportion is:
The naive calculation reports 23.3%, rather than the correct patient-level 20.0%. The numerator increased by four because four outcome patients were counted twice; the denominator increased by ten. If instead more non-outcome patients than outcome patients had been duplicated, the naive proportion could be lower than the patient-level proportion. Duplicate records do not have a fixed direction of bias: their effect depends on whose records are duplicated and how outcomes are counted.
The correction is not simply to delete any row that looks repeated. The records must first be judged to represent the same patient, and the study must define the analysis unit. If the outcome is a patient-level indicator, count each included patient once. If the intended outcome is recurrent events, preserve the event structure and avoid treating a copied record as an additional event.
Measuring and Diagnosing Linkage Error
When a set of record pairs has been validated against a more reliable identity source or careful review, linkage performance can be assessed against that reference. Sensitivity is the proportion of true matching pairs that the procedure links. Positive predictive value is the proportion of linked pairs that truly match. Sensitivity reveals missed matches; positive predictive value reveals false matches among accepted links. Both matter: a method can capture most true matches while also linking many nonmatches, or make few false links while missing many true links.
Linkage should be inspected at both the pair and cohort levels. Useful diagnostics include the distribution of match scores, the number of pairs near the decision threshold, the number of records linked to multiple people or incompatible identities, and the counts of patients and outcome events before and after deduplication. If identifiers are expected to be stable, systematic disagreements on those fields warrant investigation. Large or unexpectedly uneven changes in outcome counts can signal a linkage or counting problem, although they do not by themselves prove one.
Validation should resemble the records to which the linkage procedure will be applied. For example, a validation set composed mostly of complete records may overstate performance when the actual cohort contains many missing identifiers. Record the source and coverage of validation, and report that linkage accuracy cannot be established from the agreement pattern alone unless there is a credible reference or a justified model.
Worked Example: Interpreting Linkage Validation
In a hypothetical validation sample, review identifies 100 pairs that truly refer to the same person and 200 pairs that refer to different people. The linkage rule accepts 90 of the true matches and also accepts 10 of the nonmatches. It rejects the remaining 10 true matches and 190 nonmatches.
Sensitivity is \(90/100=0.900\), or 90.0%: the procedure links 90% of the true matching pairs in this validation sample. Positive predictive value is \(90/(90+10)=0.900\), also 90.0%: 90% of accepted pairs are true matches. These are different quantities even though they happen to be equal here. There are 10 false matches and 10 missed matches in the reviewed sample.
These results describe performance in the validation sample under its pair selection and reference standard. They do not establish that 90% of all linked pairs in a different cohort are correct. In particular, positive predictive value can change when the proportion of true matches among candidate pairs changes.
Accounting for Ambiguous Matches
Hard thresholding turns uncertain identity into a definite decision. When uncertainty is material, a useful alternative is to evaluate more than one plausible linkage configuration, weighting each by its probability under the linkage model. This carries linkage uncertainty into the patient count and outcome summary instead of pretending every borderline decision is certain. The approach depends on calibrated probabilities and a coherent representation of possible configurations. Pairwise probabilities cannot always be treated as independent: for example, linking one record to two different people may be impossible under a one-to-one rule.
Worked Example: Averaging Over Two Ambiguous Links
Consider a hypothetical cohort with 40 confirmed unique patients, including 5 with the outcome. Two additional records are each possibly duplicates: the first has match probability 0.80 and is an outcome record; the second has match probability 0.30 and is a non-outcome record. For this small illustration, assume the two linkage decisions are independent and that each unmatched record adds one new patient.
If both records are duplicates, the patient-level proportion is \(5/40=0.1250\). If only the first is a duplicate, it is \(5/41=0.1220\). If only the second is a duplicate, it is \(6/41=0.1463\). If neither is a duplicate, it is \(6/42=0.1429\). The probabilities of those four configurations are, respectively, \(0.80\times0.30=0.24\), \(0.80\times0.70=0.56\), \(0.20\times0.30=0.06\), and \(0.20\times0.70=0.14\).
Weighting each configuration-specific proportion by its probability gives an average proportion of:
Thus, the average across these four configurations is approximately 12.71%. This is not automatically the same as the proportion formed by dividing the expected outcome count by the expected patient count. It is an average of configuration-specific proportions, and the distinction matters because a proportion is a nonlinear function of its counts. In a real application, the linkage configurations and their probabilities must respect the structure of the data and the model.
Choosing an Approach and Recognizing Failure
| Approach | Useful when | Main limitation |
|---|---|---|
| Deterministic linkage | A validated unique identifier is available and recorded consistently. | Small errors or missing identifiers can create missed matches; identifier validity still needs checking. |
| Probabilistic linkage | Identifiers are incomplete or inconsistent, but several fields jointly provide evidence. | Results depend on the comparison model, priors, calibration, and decision rules. |
| Clerical review | A manageable set of ambiguous or high-impact pairs needs adjudication. | Review can be costly and subjective; decisions should be documented and, where feasible, made without outcome-driven selection. |
Common failure modes include using identifiers as though they were error-free, applying a threshold without inspecting borderline pairs, multiplying evidence from strongly dependent fields as if independent, and assuming that two records with similar values must refer to the same person. Another important mistake is to use outcomes to decide whether an uncertain pair should be merged. If linkage decisions are influenced by whether a merge makes outcome counts more plausible or results more favorable, the analysis can become outcome-dependent. Prefer identity evidence and prespecified rules; if outcomes are used in a special linkage model, the procedure and its implications require explicit justification and validation.
Linkage can improve the construction of a patient-level cohort, but it cannot by itself make external controls comparable to trial participants or establish that an observed difference is causal. It also cannot repair poor identifier quality without assumptions, nor eliminate uncertainty when multiple identity assignments remain plausible. The linkage decisions, quality checks, and resulting counts should be reported so readers can assess how the cohort was built.
Check Your Understanding
Answer each question, then compare your reasoning with the solutions below.
- In a validation set, 80 true matching pairs are present and the linkage method accepts 72 of them. What is the sensitivity, and what does it mean?
- A record pair has prior match probability 0.02 and observed evidence with likelihood ratio 10. Calculate the posterior match probability.
- Why can duplicate records increase or decrease an observed patient-level outcome proportion?
- When might it be inappropriate to collapse two enrollment rows into one patient-level row?
Solutions. 1. Sensitivity is \(72/80=0.900\), or 90.0%; the procedure linked 90% of true matching pairs in that validation set. 2. Prior odds are \(0.02/0.98=0.02041\); posterior odds are \(0.02041\times10=0.2041\), giving probability \(0.2041/(1+0.2041)\approx0.1695\), or 16.95%. 3. The direction depends on whether outcome patients or non-outcome patients are disproportionately duplicated, because duplicates can alter both the numerator and denominator. 4. If the rows represent distinct episodes that the study intends to analyze, collapsing them would remove part of the intended episode-level information; the analysis unit and episode rules must govern the decision.