Why a Seamless Trial Needs a Stable Target
A seamless multi-stage trial may move from an exploratory stage into a confirmatory stage without a break in recruitment or treatment. The operational transition can change: the sample size may increase, an arm may be dropped, or eligibility and assessment procedures may be revised. The statistical question is whether the stages still address the same treatment effect. If not, combining their results can blur distinct scientific questions, even when the protocol gives both stages the same endpoint label.
The audit is about the estimand, not merely the estimator, test, or data set. An estimand states the treatment effect of interest in a defined population, for specified treatment conditions and outcome, using a specified approach to intercurrent events and a specified population-level summary. Earlier material on separating the estimand, estimator, and test makes this distinction important: a calculation can be performed on stage data without that calculation necessarily estimating the intended common target.
The notation emphasizes that an estimand is more than a number or endpoint name. For a mean difference, for example, \(g_s\) could subtract the mean outcome under \(B_s\) from the mean under \(A_s\). The same subtraction can nevertheless yield different targets if the population, treatment conditions, outcome definition, or intercurrent-event strategy changes. Preservation requires checking the full specification, not just the summary formula.
A Stage-by-Stage Preservation Audit
Let \(\mathcal{E}_1,\ldots,\mathcal{E}_S\) describe the stages. The strongest and clearest form of preservation is that every component is the same across stages, so that \(\theta_1=\cdots=\theta_S=\theta\) refers to one common target. The stage procedures do not have to be identical: recruitment can proceed at different rates, and sample sizes can differ, provided those operational changes do not redefine the target.
Compare the eligibility rule and the population to which the effect is meant to apply. Ask whether a new stage includes a different disease severity, biomarker group, or other subgroup, and whether any subgroup effects are being averaged over the same target composition.
Record what is assigned in each condition, including dose, duration, permitted concomitant care, and the comparator. A shared treatment name does not guarantee that the treatment condition is unchanged.
Check the variable, measurement method, timing, and rules for deriving the outcome. A measure at week 12 and the same-named measure at week 16 are not automatically the same outcome target.
Check whether events such as rescue medication, treatment discontinuation, or switching are handled in the same way in both stages. If the strategy changes, the resulting outcome distributions can describe different treatment effects.
Confirm that the stages use the same population-level contrast, such as a mean difference or risk difference, and the same target weighting where a marginal effect is intended.
This checklist is deliberately component-based. A protocol amendment may leave four components untouched but change the fifth. That is enough to require a careful decision: either show that the change leaves the target parameter unchanged, define a common target and obtain stage-specific estimates of it, or acknowledge that the stages address different estimands. Merely pooling records does not resolve the mismatch.
Worked Example: Mapping a Phase Transition With an Unchanged Target
Consider a hypothetical seamless trial of a new oral therapy versus placebo for adults meeting one fixed eligibility rule. Stage 1 enrolls 120 participants and Stage 2 enrolls 240. Both stages compare assignment to the same 20-mg daily regimen for 12 weeks with assignment to placebo. The outcome is change from baseline in a symptom score at week 12, in points; rescue medication is permitted, and the treatment-policy strategy uses each participant’s week-12 outcome regardless of rescue use. The target is the mean difference in change between assigned groups in the eligible adult population.
The population is the same because the eligibility rule and target population are unchanged. The treatment conditions are the same because dose, duration, comparator, and permitted care match. The outcome is the same because the score, change definition, and assessment time match. The rescue-medication strategy is the same, and the summary is the same mean difference. Thus, the stage specifications agree on all five components, even though the number enrolled differs.
Suppose, hypothetically, the Stage 1 estimated mean changes are \(-6.0\) points for therapy and \(-2.0\) points for placebo. Its estimated contrast is \( \hat{\theta}_1=-6.0-(-2.0)=-4.0 \) points. Suppose Stage 2 estimates are \(-5.5\) and \(-1.5\) points, giving \( \hat{\theta}_2=-5.5-(-1.5)=-4.0 \) points. Both estimates target the same mean difference, \(\theta\). The matching values are illustrative; preservation follows from the matching specifications, not from the estimates happening to agree.
When the Population Mix Changes
A particularly subtle transition is one that keeps subgroup eligibility but changes the proportions of subgroups enrolled. The earlier material on population-composition effects establishes why the marginal effect can then shift even when subgroup-specific effects do not. A common target can still be specified: define the target subgroup proportions in advance and standardize each stage’s subgroup effects to those proportions.
For two subgroups, let the target proportion in subgroup 1 be \(q\), with \(1-q\) in subgroup 2. If \(\theta_{s1}\) and \(\theta_{s2}\) are stage \(s\)’s subgroup mean differences, the target-standardized stage effect is \( \theta_s^{*}=q\theta_{s1}+(1-q)\theta_{s2} \). This is a marginal mean difference for the stated target mix, not necessarily the unadjusted contrast in the stage’s actual enrollment mix. The target proportions must be justified by the scientific question; selecting them after seeing stage results can change the target rather than preserve it.
Worked Example: Standardizing Two Stages to One Target Mix
In a hypothetical trial, a lower symptom score is better. The target population is defined as 60% participants with mild disease and 40% with severe disease. Within each subgroup, the treatment contrast is the mean change under active treatment minus the mean change under control, measured at the same visit using the same intercurrent-event strategy.
Stage 1 has subgroup mean changes of \(-4\) versus \(0\) points for mild disease, and \(-10\) versus \(0\) points for severe disease. Therefore its subgroup contrasts are \(-4\) and \(-10\) points. Standardizing to the fixed target mix gives \( \theta_1^{*}=0.60(-4)+0.40(-10)=-2.4-4.0=-6.4 \) points.
Stage 2 has subgroup mean changes of \(-3\) versus \(0\) points for mild disease, and \(-11\) versus \(0\) points for severe disease. Its subgroup contrasts are \(-3\) and \(-11\) points, so \( \theta_2^{*}=0.60(-3)+0.40(-11)=-1.8-4.4=-6.2 \) points.
Both standardized estimates address the mean difference for the same 60%/40% target population, even if the stages recruited different subgroup proportions. Their numerical difference is not evidence of an estimand mismatch: sampling variation or real stage differences may explain it. By contrast, comparing unstandardized stage effects would compare effects averaged over each stage’s actual composition, which may be different targets.
Why Component Equality Justifies a Common Target
The core justification is direct. The target parameter is a function of the population, treatment conditions, outcome, intercurrent-event strategy, and summary. If those inputs are identical across stages, then the distributions and summary defining the parameter refer to the same scientific question. This conclusion is about the target; it does not establish that any particular combined estimator or test has valid operating characteristics.
A limited estimator justification follows in the simplest setting. Suppose \(K\) stage-specific estimates \(\hat{\theta}_s\) are unbiased for the same fixed \(\theta\), and \(w_s\) are fixed weights that sum to one. Then linearity of expectation gives
This short derivation shows why common-target stage estimates can be combined in this simple case. It does not show that data-dependent weights, interim selection, or a particular confirmatory analysis retain the same properties. If an adaptation selects stages, arms, populations, or analysis rules using accumulating data, the combined inference requires its own justification. Estimand preservation is necessary for a common scientific interpretation, but it is not by itself a guarantee of valid inference.
Worked Example: A Follow-Up Change Breaks the Shared Target
Suppose a hypothetical Stage 1 compares treatment and control on change in a symptom score from baseline to week 12. The observed mean changes are \(-5\) and \(-1\) points, so the estimated Stage 1 contrast is \( -5-(-1)=-4 \) points at week 12. In Stage 2, the score is measured at week 16 instead; the observed mean changes are \(-7\) and \(-1\) points, giving \( -7-(-1)=-6 \) points at week 16.
Both calculations are mean differences in symptom-score points, but the outcome component includes its assessment time. The first estimate addresses a week-12 mean difference and the second a week-16 mean difference. Their simple average, \( (-4+(-6))/2=-5 \) points, is arithmetic, not an estimate of either specified time-point effect. A common target would require choosing a time point or defining another scientifically justified outcome and estimating it comparably in both stages. Without that reconciliation, the stages do not preserve one common estimand.
Diagnostics, Alternatives, and Limits
A practical diagnostic is to write one row per stage and one column per estimand component. Record the actual protocol language and implementation, not just shorthand such as “same endpoint” or “same treatment.” Then flag each difference and ask whether it changes the target parameter. This is especially useful for amendments: small wording changes can imply a new eligibility population, treatment strategy, or assessment time.
| Finding at transition | Interpretation | Possible response |
|---|---|---|
| All five components match | Stages specify one target estimand | Use stage estimates for that target, with a separately justified analysis |
| Actual subgroup mix differs, but target mix is fixed | Raw marginal effects may target different populations | Estimate subgroup effects and standardize to the declared common mix, if data support it |
| Outcome timing or definition changes | Stages generally target different outcome quantities | Align measurement or define and estimate a defensible common outcome |
| Treatment or event strategy changes | The contrast may no longer represent the same treatment effect | Restore the common strategy, or report distinct estimands rather than treating them as interchangeable |
There are two main alternatives when exact preservation is not feasible. One is to revise procedures so both stages measure the same target. The other is to define separate stage-specific estimands and report them as such, possibly alongside an analysis that translates each to a common target when the required data and assumptions are available. The second approach can be scientifically informative, but it is not equivalent to claiming that the original stages already shared one target.
Standardization also has limits. It requires relevant subgroup-specific information and a target composition that is meaningful for the intended population. Sparse or absent data in a target subgroup can make the common-target estimate unstable or unsupported. Similarly, relabeling an outcome after the fact does not fix incompatible measurement times, and statistical adjustment cannot automatically repair a changed treatment condition or intercurrent-event strategy. A discrepancy should be resolved substantively, not concealed by pooling.
Common Mistake
Correction: The score’s timing and construction, target population, treatment conditions, intercurrent-event handling, and target weighting also define the effect. Compare those components explicitly. Even after confirming a common estimand, justify the estimator and inference separately, particularly when an adaptive decision affected which data or stages contribute.
Check Your Understanding
For each question, identify the relevant estimand components and state whether the stages automatically share one target.
- Two stages use the same eligibility rule, treatments, week-12 outcome, rescue strategy, and mean-difference summary, but enroll different numbers of participants. Is the estimand preserved?
- Two stages include the same two disease-severity groups, but their actual subgroup proportions differ. What additional specification can define a common marginal target?
- Stage 1 measures a score at week 12 and Stage 2 at week 16. Is averaging the two estimated mean differences enough to preserve a common target?
- Both stages use the same endpoint and visit, but Stage 2 changes the comparator dose. Which component must be reconsidered?
Solutions. (1) Yes, on the stated facts: sample size is not itself one of the five estimand components. The common target does not, by itself, settle whether a combined analysis is valid. (2) Specify a fixed target subgroup mix and estimate each stage’s effects standardized to that mix, if the necessary subgroup data support the calculation. (3) No. Assessment time is part of the outcome definition; an average is not automatically an estimate of either time-specific effect. (4) The treatment conditions have changed, so the treatment contrast may be different even though the endpoint and visit match.