Why Endpoint Compatibility Matters
A seamless phase 2/3 trial carries a clinical question across a stage boundary. The treatment-selection or development decision may change, but the outcome used to judge benefit still needs a clear relationship from one stage to the next. If phase 2 measures symptom improvement at week 4 while phase 3 measures hospitalization by week 24, the data do not become measurements of one outcome merely because they came from the same trial.
Endpoint compatibility is a prerequisite for interpreting evidence across stages, not a guarantee that every form of combined analysis is valid. The earlier material on preserving the estimand across the phase boundary addresses the broader target of estimation. Here the focus is narrower: whether the endpoints themselves represent the same outcome, or can be defensibly translated to a common measurement.
What Must Match Across Stages?
Start with the endpoint’s clinical meaning, then inspect how that meaning is measured. A matching label such as “response” is not enough: response might mean a 30% improvement in one stage and a 50% improvement in the other. Conversely, different instrument names do not automatically imply different endpoints if a suitable crosswalk has been established and the underlying construct and assessment conditions agree.
- Outcome construct: Both stages should measure the same clinical concept, such as pain intensity, symptom response, or time to a specified event.
- Operational rule: The event definition, response threshold, scoring algorithm, and direction of benefit must agree, or be connected by a justified mapping.
- Assessment time: The time point or window must correspond to the same clinical horizon. A week-4 outcome and a week-12 outcome are not interchangeable just because both concern symptoms.
- Intercurrent events and missing data: The rules for events such as rescue medication, treatment discontinuation, or an unobserved assessment must not silently change what the endpoint means.
- Target population and summary: Endpoint alignment does not by itself ensure that the stages address the same target population or effect summary. Those features also matter when interpreting a common estimand.
These checks should be made before looking for a convenient way to combine stage data. A post hoc redefinition to make two outcomes look alike can change the clinical question and can make the evidence difficult to interpret. Endpoint rules should therefore be written precisely in the protocol and in any transition plan.
A Justification for Requiring a Common Outcome
Suppose each stage produces an effect estimate \(\hat{\theta}_s\), where \(s=2,3\). For the moment, assume each estimate is unbiased for its stage-specific effect \(\theta_s\). A weighted summary with prespecified fixed weight \(w\), where \(0\leq w\leq1\), has expectation:
If both stages estimate the same well-defined effect \(\theta\), this expectation is \(\theta\). If the endpoint definitions differ, however, \(\theta_2\) and \(\theta_3\) may refer to different outcomes; the weighted expression then has no automatic interpretation as the effect on either one. Even when the endpoint is shared, stage-specific effects can differ because populations or other conditions differ. A common endpoint is therefore necessary for a common-outcome interpretation, but not sufficient to establish a common effect or validate a particular inferential procedure.
A scale conversion can address a measurement-unit difference, but only if its assumptions are credible. For a linear conversion \(Z_s=a_s+b_sY_s\), the conversion must preserve the relevant clinical construct and direction, and its parameters must be supported independently of an outcome-driven choice. Thresholded outcomes require particular care: converting a continuous score does not necessarily make two different responder thresholds equivalent.
Worked Example: Identical Binary Responder Endpoints
Hypothetical data. In each stage of a trial, participants are randomized to treatment or control. The endpoint in both stages is “at least a 30% reduction from baseline symptom score at week 12.” Assume the same scoring instrument, assessment window, rescue-medication rule, and missing-outcome rule are used in both stages. Each arm has 100 participants with an observed endpoint under the stated analysis rule.
In phase 2, 50 of 100 treatment participants and 30 of 100 control participants respond. The risk difference, defined as treatment response proportion minus control response proportion, is:
In phase 3, 44 of 100 treatment participants and 32 of 100 control participants respond:
Each difference is expressed as a proportion of week-12 responders under the stated rule (equivalent to 20 and 12 percentage points, respectively). Because the endpoint definitions and assessment rules agree, these stage estimates concern the same kind of outcome. Their different numerical values do not by themselves indicate an endpoint incompatibility. They also do not prove that a pooled analysis is appropriate: population, treatment, and inferential design considerations still need examination.
When an Endpoint Change Calls for Separate Evidence
A changed endpoint calls for separate stage-specific evidence when the two outcomes cannot be interpreted as measurements of one prespecified clinical target. Examples include changing from a continuous symptom score to a clinical event, changing the assessment horizon in a way that changes the question, or adopting a new response threshold without a defensible bridge. In these circumstances, report and interpret evidence for each endpoint on its own terms. Do not present a single pooled effect as if it estimated one shared endpoint.
“Separate evidence” does not mean that the earlier stage is irrelevant. It means that its contribution must be described according to the endpoint it actually measured—for example, as supportive evidence about an earlier outcome rather than as direct observations of a later primary outcome. Whether and how such evidence can enter a confirmatory test is a distinct inferential question; endpoint similarity alone cannot answer it.
Worked Example: A Symptom Score Changes to a Responder Endpoint
Hypothetical data. Phase 2 measures mean change in a symptom score from baseline to week 4, with lower scores indicating improvement. The mean changes are \(-4.1\) points for treatment and \(-0.9\) points for control. The treatment-minus-control contrast is:
Phase 3 instead measures the proportion with at least a 30% improvement at week 12. Suppose 44 of 100 treatment participants and 32 of 100 control participants meet that rule. The treatment-minus-control difference is:
or a difference of 12 percentage points. The phase 2 result is a mean contrast in week-4 score change, measured in points; the phase 3 result is a week-12 difference in responder proportions. They differ in outcome scale, summary, and time horizon. Adding \(-3.2\) points to \(0.120\), or averaging these numbers, has no meaningful interpretation. The appropriate conclusion is that each stage provides evidence about its own endpoint; the phase 2 estimate is not a direct estimate of the phase 3 responder effect.
Harmonization: When Different Instruments Can Still Align
An instrument change may be manageable if a prospective validation study establishes that the measurements represent the same construct and provides a reliable conversion over the range relevant to the trial. The mapping should be specified before it is used to support the cross-stage interpretation. A simple multiplication or change of units is not enough if the instruments differ in content, sensitivity, or the way participants use their response scales.
Worked Example: Converting a Validated Pain Scale
Hypothetical data. Phase 2 uses a 0–10 pain scale and phase 3 uses a 0–100 scale. Before the trial, an independent validation establishes that, for this purpose and range, the phase 3 scale can be converted by dividing by 10. Both stages assess mean change from baseline to week 8, with negative changes indicating less pain. Assume the same treatment contrast and missing-measurement strategy are intended.
In phase 2, mean changes are \(-2.4\) points with treatment and \(-0.9\) points with control. The treatment-minus-control contrast is \(-1.5\) points on the 0–10 scale.
In phase 3, mean changes on the 0–100 scale are \(-21\) points with treatment and \(-7\) points with control. First calculate the contrast on that scale, then convert:
Under the stated validation and matching endpoint rules, the phase 3 contrast is interpretable on the phase 2 scale, and the two estimates address a comparable week-8 pain-change outcome. The converted estimate is \(-1.4\) points, not \(-14\) points on the 0–10 scale. This harmonization supports endpoint comparability; it does not establish that the stages can automatically be pooled or that their treatment effects are identical.
Practical Checks and Alternatives
A useful endpoint review asks whether an independent reader could apply the same clinical definition to either stage and understand the resulting estimate in the same units and time frame. If not, specify the discrepancy and its consequence before interpreting combined evidence.
| Situation | Endpoint assessment | Reasonable response |
|---|---|---|
| Same outcome rule, scale, time point, and handling rules | Directly comparable endpoint definitions | Evaluate whether other design and inferential conditions also support combined interpretation |
| Different instruments, with a prespecified and independently validated mapping | Potentially harmonizable | Apply the validated conversion and examine whether its assumptions fit the observed range |
| Different outcome construct, threshold, or clinically important time horizon | Not a common endpoint without a defensible bridge | Keep evidence stage-specific, or prospectively redesign the endpoint before relying on cross-stage interpretation |
Warning signs include inconsistent endpoint language across the protocol and analysis plan; thresholds that differ between stages; assessment windows that overlap only partially; a conversion derived after examining treatment differences; and substantial missingness handled differently across stages. These do not all prove that comparison is impossible, but each requires resolution rather than assumption.
A common mistake is to reason, “Both outcomes are measures of the same disease, so their results can be pooled.” That is too broad. Disease area is not an endpoint definition: a symptom score, a responder indicator, and a hospitalization event can all concern the same disease while answering different clinical questions. The correction is to compare the actual variable, rule, time point, and summary, and to require a justified mapping before treating distinct measurements as one endpoint.
Check Your Understanding
Answer each question, then compare your reasoning with the solutions.
- Phase 2 defines response as a 20% improvement at week 8; phase 3 defines it as a 40% improvement at week 8. Are these automatically the same endpoint? What should be done with their estimates?
- A validated conversion divides a phase 3 score by 10 to express it on the phase 2 scale. The phase 3 treatment-control contrast is \(-16\) points before conversion. What is the converted contrast?
- In phase 2, the risk difference is \(0.18\); in phase 3, it is \(0.10\). State each in percentage points and explain what endpoint information is needed before judging comparability.
- Why is a weighted average of a week-4 mean symptom-score contrast and a week-12 event risk difference not automatically a meaningful treatment effect?
Solutions. (1) No. The thresholds define different responder outcomes even though the time point agrees. Unless a prospectively justified bridge supports a common rule, interpret the stage estimates separately. (2) \(-16/10=-1.6\) points on the phase 2 scale. The conversion is interpretable only under the stated validation assumptions. (3) The differences are 18 and 10 percentage points, respectively. Compare the outcome rule, assessment window, missing-data and event-handling rules, and target summary; the numbers alone do not establish endpoint compatibility. (4) The component estimates measure different outcomes, use different units, and refer to different horizons. Their weighted average has no automatic clinical interpretation as an effect on either endpoint.