From a Clinical Question to a Complete Specification
An estimand is a precise description of the treatment effect a trial seeks to quantify. A question such as “Does the new therapy help?” is not yet precise enough: the answer can depend on who is considered, which treatment conditions are compared, what outcome and time point are used, how events after randomization are handled, and how outcomes are summarized across people. The five attributes make those choices explicit.
An intercurrent event is an event after treatment initiation that affects either the interpretation or the existence of the outcome of interest. Examples include stopping assigned treatment, starting rescue medication, or switching treatment. Such an event is not automatically the same thing as a missing outcome: the event may occur even when an outcome is later observed. For a complete estimand, state a strategy for each relevant event rather than leaving its consequences implicit.
These attributes are connected. For example, the outcome and event strategies together define which outcome value is relevant after an event; the population-level summary then states how those values are combined and compared. Naming only the treatment arms and an endpoint does not complete the specification.
What Each Attribute Must Say
Target population. Identify the people to whom the treatment effect refers. This can be defined by eligibility criteria, disease severity, setting, or another clinically meaningful description. “Trial participants” may be insufficient if the intended population is narrower or broader than the group enrolled. The target population should be stated independently of which participants happen to provide complete outcome data.
Treatment conditions. Name each condition being compared, including relevant features such as dose, route, duration, and background care. “Drug versus control” may be ambiguous if there are several doses or if background care differs. In a randomized trial, the conditions often refer to assignment to specified regimens; the attribute should make clear what those regimens are.
Outcome variable. State what is measured, how it is represented, and when it is assessed. Specify, for example, whether the endpoint is a score, a change from baseline, a binary response, or time to an event. A direction of benefit is useful where the scale has an interpretable direction. “Clinical response” without a definition or assessment time is not a sufficiently defined variable.
Strategy for each intercurrent event. List the events that matter to the clinical question and state how the outcome is defined in relation to each one. A strategy is part of the target question, not an after-the-fact convenience for handling inconvenient observations. One trial may need different strategies for different events. The later material on individual intercurrent-event strategies examines their interpretations in depth; here the task is to state the chosen strategy as one of the five attributes.
Population-level summary. Choose the quantity that summarizes the outcome in the target population and the comparison between treatment conditions. Examples include a difference in means, a difference in response probabilities, or a ratio of probabilities. Specify the contrast order and scale. A statement such as “compare outcomes” does not reveal whether the target is a mean difference, a risk difference, or another summary.
Worked Specifications
Worked Example: Pain Change With Discontinuation and Rescue Medication
Hypothetical trial. Adults with moderate-to-severe chronic pain are assigned to a new oral therapy or placebo, both with standard care. Pain is recorded on a 0-to-10 scale at baseline and week 12, with higher scores indicating worse pain. Participants may discontinue assigned treatment or begin rescue medication.
Five attributes. The target population is adults meeting the stated chronic-pain and severity criteria. The treatment conditions are assignment to the specified oral therapy plus standard care and assignment to placebo plus standard care. The outcome is change in pain score from baseline to week 12, in scale points. For both discontinuation and rescue medication, choose a treatment-policy strategy: the week-12 outcome remains the outcome of interest regardless of whether the event occurred, and the comparison is by assigned condition. The population summary is the mean change under the therapy condition minus the mean change under the placebo condition.
To make the contrast concrete, suppose the hypothetical target-population mean change is \(-2.4\) points under therapy and \(-1.5\) points under placebo. Then
Because higher pain scores are worse, a more negative change represents greater improvement. The hypothetical contrast of \(-0.9\) points therefore favors therapy in direction: the mean reduction is 0.9 points greater under therapy. This sign alone does not establish statistical or clinical significance. The calculation illustrates the defined population summary; it is not a claim about observed trial data.
Worked Example: A Response Outcome That Includes Death
Hypothetical trial. Adults with a serious condition are assigned to a new treatment or standard care. At week 24, the outcome is a binary response: alive at week 24 and improved by at least 5 points on a prespecified symptom scale. Death before week 24 makes the response outcome equal to “No.” Discontinuation of assigned treatment is handled with a treatment-policy strategy, so response is assessed regardless of discontinuation.
The target population is adults meeting the trial’s eligibility criteria. The treatment conditions are assignment to the specified new treatment and assignment to standard care. The outcome is the defined week-24 response indicator. The strategy for death is a composite strategy, because death is incorporated into the outcome as nonresponse; the strategy for discontinuation is treatment-policy. The population summary is the response-probability difference, new treatment minus standard care.
Suppose, hypothetically, that the target-population response probabilities under the two conditions are 0.62 and 0.48. The specified summary is
or 14 percentage points. Under this definition, the new treatment condition has a 14-percentage-point higher probability of meeting the response outcome. The wording matters: this response outcome includes both survival and the symptom improvement threshold, so the contrast is not a symptom-score difference among survivors.
Worked Example: Defining an Outcome in the Absence of Rescue Medication
Hypothetical trial. Adults with hypertension are assigned to a new medicine or an active comparator. Systolic blood pressure is measured at six months. Some participants may start rescue medication, while others may discontinue their assigned medicine. Suppose the clinical question concerns blood pressure at six months in a scenario in which rescue medication was not started; discontinuation is handled with a treatment-policy strategy.
The target population is eligible adults with hypertension. The treatment conditions are assignment to the specified new medicine and assignment to the specified active comparator. The outcome variable is six-month systolic blood pressure in mm Hg. The strategy for rescue medication is hypothetical: the target outcome is the six-month blood pressure that would apply in the scenario where rescue medication was not started. The strategy for discontinuation is treatment-policy. The population summary is the mean systolic blood pressure under the new-medicine condition minus the mean under the comparator condition.
For illustration, suppose the hypothetical target-population means under those definitions are 128 mm Hg and 134 mm Hg. Then
The negative contrast means the target mean is 6 mm Hg lower under the new-medicine condition than under the comparator condition. It describes the specified hypothetical scenario, not necessarily the contrast that would be obtained by simply analyzing observed blood pressures among participants who did not use rescue medication. Whether the data and analysis can support the target quantity is a separate question.
Why the Attributes Need to Agree
A short derivation shows how the attributes define a population contrast. Let \(Y^{a,\mathbf{s}}\) denote the outcome under treatment condition \(a\), defined using the specified strategies \(\mathbf{s}\) for the relevant intercurrent events. Let \(\mu_a^{(\mathbf{s})}\) be its mean in the target population. If the chosen population-level summary is a mean difference for conditions \(A\) and \(C\), the target quantity is
The equation is a direct consequence of the definition of a difference in population means: take the mean outcome defined for condition \(A\), then subtract the mean outcome defined for condition \(C\). The population attribute determines where those means are taken; the treatment conditions determine \(A\) and \(C\); the outcome and event strategies define \(Y\); and the summary attribute specifies subtraction in this order. If any of these changes, the target quantity can change.
For example, reversing the subtraction changes the sign but not the underlying pair of means. Changing a response probability difference to a risk ratio changes the scale and interpretation. Changing the event strategy can change which outcomes the population means represent. Therefore, a numerical contrast is not fully interpretable unless its five attributes are known.
Diagnostics, Alternatives, and Limitations
A useful diagnostic is to read the specification aloud as one sentence: “In [target population], compare [treatment conditions] on [outcome and time], using [strategy] for each event, summarized by [population-level contrast].” If a phrase is missing or could describe multiple choices, the estimand is not yet sufficiently precise. Another check is to list every relevant event and verify that each has an explicit strategy; an event should not be silently omitted merely because its outcome is often observed.
Common warning signs include an endpoint with no assessment time, treatment arms with no dose or background-care details, an event listed without its strategy, and a summary that omits the contrast order or scale. Also check coherence: a specified strategy and outcome must jointly define what value contributes after the event. If death is counted as nonresponse, for instance, the outcome is not simply the symptom score at week 24 for every participant.
Alternative estimands can answer different legitimate clinical questions. The same trial can be described using a treatment-policy strategy for an event, a hypothetical strategy, or another strategy, but these are not interchangeable labels for one unchanged question. Likewise, a mean difference and a response-probability difference summarize different features of outcomes. The appropriate specification depends on the clinical question, not on which version produces a more favorable result.
A complete estimand specification does not guarantee that the trial data identify the target quantity, that an analysis estimates it without bias, or that the outcome is measured without error. Those issues require separate assessment. Nor does specifying an estimand determine the statistical model, missing-data assumptions, or uncertainty calculation. The five attributes define what is being targeted; they do not by themselves settle how reliably it can be learned from observed data.
Check Your Understanding
For each question, identify the missing or relevant attributes. Short solutions follow so that the reasoning can be checked.
- A protocol says, “Compare the new drug with control on symptom improvement.” Name at least three details needed to make the estimand specification complete.
- A response is defined as being alive and having a symptom reduction of at least 4 points at week 16. How does the definition handle death before week 16?
- Suppose mean changes are \(-3.1\) points under treatment and \(-2.2\) points under control, and lower change indicates greater symptom improvement. Calculate the treatment-minus-control mean difference and state its direction.
- Why is “the mean outcome among participants who did not use rescue medication” not automatically the same target as an outcome defined under a hypothetical no-rescue strategy?
Solutions. (1) Specify the target population; the exact treatment conditions, including relevant regimen details; the symptom variable and assessment time; a strategy for each relevant intercurrent event; and the population-level summary with contrast order. (2) Death is incorporated into the composite response definition: a person who dies before week 16 cannot meet the stated response. (3) The difference is \((-3.1)-(-2.2)=-0.9\) points. Since lower change indicates greater improvement, the direction favors treatment; the sign alone does not establish statistical or clinical significance. (4) The observed group that did not use rescue medication is selected according to what happened after assignment. A hypothetical no-rescue strategy instead defines the target outcome in a scenario where rescue medication was not started; the two quantities need not refer to the same population or outcomes.