What Must Stay the Same for a Target to Be Preserved?
An adaptive trial can change its sample size, randomization probabilities, or other features of its conduct as information accumulates. Those changes do not automatically change the treatment effect the trial is intended to estimate. But a trial cannot claim to preserve its target merely because it keeps the same broad scientific question. The target must be specified precisely enough to check whether the relevant features remain unchanged.
An estimand is the precise quantity a study seeks to estimate. In this setting, it describes a treatment effect by specifying five attributes: the population, treatment conditions, outcome variable, strategy for handling intercurrent events, and population-level summary. An adaptation preserves the estimand only when it leaves these attributes, as defined in the protocol, unchanged. It may still change the amount or pattern of information available to estimate that quantity.
The five attributes are connected. For example, a mean difference is not fully defined without specifying whose outcomes are averaged, which treatment conditions are compared, when and how the outcome is measured, and what happens to the definition of that outcome after an intercurrent event. The summary measure alone cannot identify the target.
Specifying the Five Attributes
The attributes should be stated in operational terms, so a reader can tell whether a proposed protocol change affects the target.
- Population: the group to which the treatment-effect question refers, described by relevant eligibility or baseline characteristics.
- Treatment conditions: the interventions being compared, described sufficiently to distinguish them. A condition may include the assigned treatment, dose, schedule, and other features relevant to what is being compared.
- Outcome variable: the variable used to assess the treatment effect, including its measurement and time point where these define the outcome.
- Intercurrent-event strategy: how the target relates to an event after treatment begins that affects either the interpretation or existence of the outcome. Examples include treatment discontinuation or use of rescue medication. The strategy must say how such events enter the treatment-effect question; simply recording that they occurred is not a strategy.
- Population-level summary: the way outcomes are summarized and compared across the population, such as a difference in means, a risk difference, or a ratio of risks.
Defining the estimand is distinct from deciding how to estimate it. A model, weighting procedure, or test is an analysis choice. It may be appropriate or inappropriate for the target, but it does not itself define the target. Similarly, the adaptive randomization rule determines how participants are assigned; it is not a substitute for specifying the treatment effect.
Worked Example: Writing a Complete Estimand
Suppose a hypothetical trial compares an investigational treatment with placebo for adults who meet a prespecified set of baseline criteria. The outcome is a symptom score at week 24, where lower scores indicate fewer symptoms. Participants may discontinue assigned treatment or start rescue medication, but the protocol intends to include their week-24 scores regardless of these events.
Population: adults who meet the trial’s stated baseline criteria. Treatment conditions: assignment to the investigational treatment or placebo, as specified in the protocol. Outcome: symptom score at week 24. Intercurrent-event strategy: a treatment-policy strategy—the week-24 outcome remains part of the target regardless of discontinuation or rescue medication. Population-level summary: the difference between the population mean scores under the two assigned conditions.
For this example, let \(Y^{a}\) denote the week-24 score under assigned condition \(a\), with outcomes interpreted according to the stated event strategy. The target is
where \(P\) denotes the specified population. If hypothetical trial data gave mean scores of 12.4 points under the investigational condition and 15.1 points under placebo, the plug-in contrast would be
The negative contrast indicates a mean score 2.7 points lower under the investigational condition in this hypothetical data set. It is an estimate, not the estimand itself. Its interpretation also depends on the outcome and event strategy being the ones specified above.
Why Conduct Changes Do Not Always Change the Estimand
Consider an adaptation that changes the randomization ratio or increases the planned sample size, while retaining the same population, treatment conditions, outcome, event strategy, and summary. The target remains the same: for example, the difference between the same two population mean outcomes. The adaptation can change precision, the allocation of observations across arms, and the procedure needed for valid inference. Those are important consequences, but they are not, by themselves, changes to the estimand.
A short justification makes the distinction clear. Let \(\mu_a=E(Y^a\mid P)\) be the population mean under condition \(a\). If a trial has stages \(k=1,\ldots,K\), and the outcome distribution for condition \(a\) in each stage represents the same target population and condition, then each stage has mean \(\mu_a\). For any nonnegative stage weights \(w_k\) summing to one, the weighted mean remains
Thus changing how many observations come from each stage does not, by itself, change the defined population mean. The corresponding treatment contrast remains \(\mu_1-\mu_0\). This is a statement about the target, not a guarantee that any particular pooled estimator is unbiased or that a conventional standard error remains valid. If stage populations differ, outcomes are affected by calendar time, or adaptation influences which observations are included, the analysis must address those features.
Worked Example: A Sample-Size Adaptation With a Preserved Target
In a hypothetical trial, the target is the mean week-24 symptom-score difference between the same two treatment conditions in the same eligible population, using the same event strategy. After an interim review, recruitment continues with a larger second stage. Assume, for illustration, that the population and outcome distributions are stable across stages, outcomes are collected as specified, and the stage samples are comparable representations of the target population.
In stage 1, 40 participants in the investigational group have a mean score of 14 points, and 40 in the control group have a mean score of 16 points. In stage 2, 60 investigational-group participants have a mean of 13 points, and 30 control-group participants have a mean of 15 points. The pooled means are
The pooled contrast is \(13.4-15.57=-2.17\) points. Increasing enrollment and changing the stage-specific allocation did not change the stated estimand. The values above are hypothetical estimates; whether this pooled calculation supports valid inference depends on the design and analysis, including the assumptions stated for this example. In particular, the unchanged target does not make adaptive sampling irrelevant to uncertainty or error control.
When an Adaptation Changes the Target
An adaptation can alter an estimand even if the trial retains the same treatment labels. Adding a newly eligible group can change the population; changing the outcome time point can change the outcome variable; changing whether outcomes after rescue medication count can change the intercurrent-event strategy. Replacing a mean difference with a risk ratio changes the population-level summary. Each change may be scientifically justified, but it should not be described as preserving the original estimand without further specification.
Worked Example: Expanded Eligibility and a Changed Population
A hypothetical trial initially targets adults with severe baseline symptoms. Suppose the treatment effect on the score is a mean difference of \(-4\) points in that population. An adaptation later allows adults with milder symptoms to enroll as well; suppose the corresponding effect in the mild group is \(-1\) point. If the final enrolled mix is 60% severe and 40% mild, a simple weighted contrast for that mix is
This \(-2.8\)-point contrast targets the stated 60:40 mix, not the original severe-symptom population’s \(-4\)-point contrast. The arithmetic is not a general recipe for correcting population changes: it illustrates that a population-level summary depends on which population is being summarized. If the trial intends to retain the original target, it must specify and justify an approach that relates the data from the expanded enrollment to that target.
A useful check is to compare the protocol’s five attributes before and after an adaptation. Ask whether the same people are targeted, the same conditions are compared, the same outcome is defined, events are handled in the same way, and the same summary is used. A difference in any answer signals either a changed estimand or a need to explain why the target is nevertheless preserved.
Assumptions, Diagnostics, and Limitations
The estimand is a definition, so it does not require a statistical model to exist. Estimating it from trial data does require assumptions and an analysis suited to the design. In a randomized trial, random assignment supports comparison of treatment groups, but adaptive assignment must still maintain a meaningful chance of receiving each relevant condition for participants in the target population. The outcome must be measured and handled in accordance with its definition, and the analysis must reflect the actual assignment and adaptation procedures. If an event strategy requires outcomes that are not observed, additional assumptions or methods may be needed; the target alone does not supply those data.
Useful diagnostics include checking that eligibility and treatment definitions were applied consistently across stages, comparing baseline characteristics and outcome collection by stage, and documenting whether the outcome time point or event handling changed. Large stage differences do not automatically prove that the estimand changed, but they can warn that a pooled estimate may not represent the stated target. Likewise, a change in allocation ratios can affect precision and inference even when all five attributes are fixed.
The five-attribute specification cannot by itself guarantee that the target is identifiable from observed data, that the chosen estimator is valid, or that the adaptive design controls its intended error rates. Nor does the label “same estimand” imply that results are comparable if treatment versions, measurement procedures, or populations changed in ways left unspecified. It provides a disciplined way to state what is being targeted; design and analysis must then support estimation of that target.
Common Mistake: Treating a Stable Treatment Label as a Stable Estimand
Wrong reasoning: “The investigational treatment and placebo are still the two arms, so the estimand is unchanged.” This overlooks the other four attributes. If eligibility expands, the outcome time changes, or post-discontinuation outcomes are handled differently, the target may have changed even though the arm names did not.
Correction: Write down all five attributes for the original and adapted trial, then compare them one by one. Also distinguish the target from the estimator: the target may be unchanged while the data structure, precision, or valid analysis changes substantially.
Check Your Understanding
Answer each question, then compare your reasoning with the brief solutions.
- A trial increases enrollment but retains all five stated attributes. Has the estimand necessarily changed?
- Name the five attributes of an estimand.
- A protocol changes the primary outcome from a score at week 12 to the same score at week 24. Which attribute has changed?
- In a hypothetical population, the mean outcome is 10 under condition A and 13 under condition B. Calculate the mean difference A minus B and state its interpretation.
- Why does an unchanged estimand not guarantee that the original analysis remains valid after an adaptive change?
Solutions: (1) No. A larger sample can change information and precision without changing the five attributes. (2) Population, treatment conditions, outcome variable, intercurrent-event strategy, and population-level summary. (3) The outcome variable’s time point has changed, so the target is not automatically the same. (4) \(10-13=-3\) outcome units; the mean outcome under A is 3 units lower than under B in the specified population. (5) Adaptation can alter assignment probabilities, stage composition, or the dependence structure of the data, so estimation and uncertainty calculations must reflect the actual design.