Why Endpoint Mapping Matters
A trial endpoint is a clinical definition; a registry, claims database, or electronic health record contains recorded data. The same phrase, such as “hospitalization for heart failure,” may refer in a trial to an event reviewed against clinical records by an adjudication committee, but in claims to a billing code, or in an electronic record to an admission entry. These are not automatically equivalent measurements. Before external outcomes can be compared with trial outcomes, the clinical event must be translated into a reproducible rule for the available data, and the consequences of any difference must be made explicit.
This matters especially when external outcomes contribute information to a Bayesian analysis. A prior distribution or borrowing procedure can account for uncertainty about parameters, but it cannot make two differently measured outcomes clinically identical. If the external endpoint is a noisy proxy, treating its event rate as though it were measured exactly like the trial endpoint can produce misleading evidence. Endpoint mapping is therefore both a clinical-definition task and a measurement task.
Specify the Clinical Outcome Before the Data Rule
Mapping should begin with the trial endpoint, not with whichever fields happen to be convenient in the external data. State the clinical event and its components, the qualifying threshold or confirmation rule, and any details that determine whether an occurrence counts. For example, a trial definition might require a new heart-failure admission, documented symptoms, and treatment intensification. A rule that counts any encounter carrying a heart-failure code may capture a broader set of events.
A useful endpoint specification distinguishes at least four elements: what event qualifies, how it is measured or confirmed, which recorded fields implement the rule, and what the rule cannot establish. The operational rule should also state how repeated records for the same clinical episode are treated and which records are required to classify an event. The earlier material on Defining the External-Control Estimand and Emulating a Target Trial With Nontrial Data establishes the target outcome and study question; endpoint mapping makes the outcome measurable in a particular source.
Do not resolve every data limitation by silently changing the clinical definition. If the trial requires a confirmed event but the source provides only an unconfirmed code, record that as a measurement gap. If the available data cannot distinguish qualifying from nonqualifying events, the endpoint may need to be labelled a proxy, or the source may be unsuitable for this outcome. A changed rule can sometimes define a defensible alternative endpoint, but it is not the same endpoint merely because it uses a similar name.
From Clinical Definition to Operational Rule
The operational definition should be specific enough that two analysts using the same source and data extract would classify records in the same way. It should name the fields and qualifying values, the logic connecting them, and any required corroboration. It should also distinguish a positive classification from an absence of evidence. For instance, no hospitalization record could mean no hospitalization, incomplete capture, or a record not available in the data source. The rule should not interpret “not recorded” as “did not occur” without a justified ascertainment assumption.
| Mapping element | Question to answer | Illustrative specification |
|---|---|---|
| Clinical event | What event does the trial count? | Admission for acute heart failure, rather than any heart-failure-related contact |
| Measurement rule | How is the event confirmed? | Trial adjudication using clinical documentation; external source may have an admission record and diagnosis field |
| Operational rule | Which data make the endpoint positive? | Admission record plus a qualifying diagnosis in a specified position |
| Known limitation | What could be missed or misclassified? | Events outside captured facilities; codes may not establish acute clinical criteria |
The table is a specification aid, not a substitute for source assessment. In particular, a diagnosis code’s presence does not by itself establish that the clinical event met the trial’s criteria. Conversely, requiring multiple pieces of documentation may improve specificity but miss events that were incompletely recorded. These trade-offs should be described in clinical terms and, where possible, evaluated empirically.
Worked Example: Translating an Adjudicated Hospitalization Endpoint
Suppose a hypothetical trial defines its endpoint as a hospitalization for acute heart failure, confirmed by review of clinical records. An external claims source records inpatient admissions and diagnosis codes but does not contain the adjudication documents. The mapping team proposes a rule requiring an inpatient admission record and a heart-failure diagnosis in the primary diagnosis position.
Clinical target. The event is an acute heart-failure hospitalization, not an outpatient visit, a historical diagnosis, or any admission for any reason among a patient with heart failure.
Operational rule. Count an episode when the source contains an inpatient admission and the specified primary diagnosis field contains a qualifying heart-failure diagnosis. The rule should specify how the source identifies an inpatient admission and how duplicate claims for one episode are handled; the diagnosis-code vocabulary itself must be specified and audited separately.
What is and is not established. An admission plus a diagnosis code is directly observed as a recorded claim pattern. It is not direct evidence that clinical reviewers would have confirmed the trial’s acute-event criteria. The external endpoint should therefore be described as a claims-defined proxy for adjudicated acute heart-failure hospitalization unless validation supports stronger equivalence.
Assumption to examine. To treat its rate as comparable to the trial rate, one would need the claims rule to identify qualifying clinical events with adequate accuracy, and the measurement process must not differ in a way that distorts the comparison. The rule alone does not establish these assumptions. A validation exercise or a sensitivity analysis is needed if misclassification could materially affect the conclusion.
Direct Events, Proxies, and Misclassification
The distinction between a direct event and a proxy is about the relationship between the recorded item and the clinical target. A registry entry for death directly records death, although dates or vital status can still be missing or erroneous. A diagnosis code is generally a proxy for the clinical diagnosis because it may reflect billing, coding practice, or a differential diagnosis rather than confirmed disease. An electronic-record laboratory result directly records a measured test result, but it is not necessarily equivalent to a trial endpoint requiring a threshold on repeated measurements or central laboratory confirmation.
For a binary clinical event, let \(E=1\) denote that the target event truly occurred under the clinical definition, and let \(R=1\) denote that the operational real-world rule classifies the event as present. The rule’s sensitivity is \(P(R=1\mid E=1)\), and its specificity is \(P(R=0\mid E=0)\). Sensitivity describes the fraction of true events captured; specificity describes the fraction of non-events correctly classified. Both are conditional on the target definition and the data system in which the rule is used.
If the true event prevalence is \(p=P(E=1)\), the observed proxy-positive proportion is \(q=P(R=1)\). By separating proxy positives among true events from proxy positives among non-events:
Writing sensitivity as \(Se\) and specificity as \(Sp\), and using \(P(R=1\mid E=0)=1-Sp\), gives:
Rearranging yields a corrected prevalence when the accuracy parameters are known and the proxy is informative:
This derivation assumes a well-defined binary target, a relevant and correctly estimated sensitivity and specificity, and a measurement process that is adequately represented by those accuracy values. In practice, \(Se\) and \(Sp\) are estimated and uncertain. The correction is not a replacement for validation, and a corrected point estimate alone does not carry that uncertainty into an analysis. Moreover, accuracy can vary across populations or data systems, so values from one setting may not apply to the external cohort.
Worked Example: Correcting a Proxy-Defined Event Rate
In a hypothetical external cohort of 200 patients, a claims rule classifies 30 as having a target event, so the observed proxy-positive proportion is \(q=30/200=0.15\). A separate hypothetical validation sample suggests sensitivity \(Se=0.80\) and specificity \(Sp=0.95\) for this rule against the clinical event definition.
Under the assumptions that these accuracy values apply to the cohort and that the binary misclassification model is appropriate:
The corrected estimated event proportion is about 13.3%, corresponding to \(200(0.1333)=26.7\), or approximately 27 events in a cohort of 200. This is not a count of individually identified patients: it is a model-based correction to the proportion. The uncorrected proxy rate is 15.0%. Here, false positives among patients without the event inflate the proxy rate, while missed true events reduce it; the specified accuracy values imply a net downward correction.
The calculation treats \(Se\) and \(Sp\) as fixed and known, even though in practice they are estimated. A full analysis should account for their uncertainty and assess whether the validation sample represents the external cohort. If accuracy differs substantially by patient subgroup or site, a single pair of values may be inappropriate.
Compare the Measurement Process, Not Just the Endpoint Name
A practical comparison asks whether trial and external data measure the same clinical construct, whether the operational rules correspond, and whether differences are likely to change event classification. A registry may directly capture death but provide less detail about the cause of death. Claims may capture billed encounters but not the clinical evidence used by trial adjudicators. An electronic record may include test values but have incomplete capture when patients receive care elsewhere. Each source can be useful, but the endpoint description must match what it actually measures.
Where evidence allows, validation can compare the operational rule with a reference assessment of the clinical event in a subset of records. The reference assessment should itself be defined clearly; it may be chart review or adjudication, not an infallible truth. Report the population and setting used for validation, the number of events and non-events assessed, and uncertainty around estimates of sensitivity and specificity. If validation is unavailable, say so rather than presenting an assumed proxy as verified.
Endpoint mapping differs from two common alternatives. One is to use the trial’s original endpoint label without adapting it to the external source; this obscures what was actually measured. The other is to choose a convenient available measure and treat it as equivalent without validation; this can change the clinical meaning while preserving a familiar name. A more defensible approach specifies the external rule, documents its relationship to the target event, and either justifies comparability or retains the distinction in interpretation and analysis.
Diagnostics, Limitations, and Common Mistakes
Warning signs include a rule that cannot be reproduced from its written specification, large differences in event rates when a reasonable operational rule is changed, an implausibly low event rate in a source expected to capture the event, or strong variation in coding or recording practices across settings. Such patterns do not prove misclassification, but they warrant investigation. Compare alternative defensible definitions where possible, inspect the records underlying classifications, and make clear whether differences reflect the clinical event or how it was recorded.
A central limitation is that proxy accuracy is often not known for the population and period of interest. A rule validated in one hospital system may perform differently in another; even within a system, changes in documentation practice may alter capture. A direct measurement can also be incomplete, delayed, or recorded with error. Endpoint mapping makes these issues visible, but it does not by itself remove them or establish causal comparability between trial and external patients.
- Mistake: “The claims code has the same name as the trial endpoint, so the outcomes match.” A code is a recorded classification; the trial endpoint may require clinical criteria and adjudication. Describe the code-based outcome as a proxy unless equivalence is supported.
- Mistake: “No record means no event.” This conclusion requires sufficiently complete capture and ascertainment. If care outside the source is not observed, no record may mean the event was not captured.
- Mistake: “A direct data field cannot be wrong.” Direct recording narrows the inference from data to event, but fields can still be missing, misdated, or erroneous. Assess the quality relevant to the endpoint.
- Mistake: “A corrected rate solves all comparability problems.” A correction depends on accuracy assumptions and does not align populations, clinical definitions, or other aspects of study design.
Check Your Understanding
Answer each question using the distinction between a clinical event and its recorded indicator.
- A trial counts death from any cause, while an external registry records vital status but has an unknown date for some deaths. Is death directly recorded or proxied, and what limitation remains?
- A claims algorithm has sensitivity 0.90 and specificity 0.90. If the true event proportion is 0.20, what proxy-positive proportion does the misclassification equation predict?
- Why is “no diagnosis code found” not automatically equivalent to “the clinical event did not occur”?
- An electronic record directly contains a laboratory result, but the trial endpoint requires two qualifying measurements. What must an operational mapping rule specify?
Solutions. (1) Vital status is a direct record of death, but an unknown date limits classification of when death occurred; the event can be observed while its timing remains uncertain. (2) Using \(q=Se\,p+(1-Sp)(1-p)\), \(q=0.90(0.20)+0.10(0.80)=0.18+0.08=0.26\). The predicted proxy-positive proportion is 26%. (3) The event may have occurred without being coded or captured in that source, so absence of the record is not evidence of absence unless capture is sufficiently complete. (4) It must specify the qualifying threshold, the requirement for two measurements, and how the available values establish that requirement; one recorded result alone does not implement a two-measure endpoint.