Why an Estimand Needs Five Attributes
A trial can report a treatment difference without making clear what treatment question that difference answers. The ambiguity may concern who the result is intended to describe, when an outcome is measured, or how events after randomization are treated. These are not merely reporting details: changing any one can change the target quantity, even if the trial data and analysis code stay the same.
The ICH E9(R1) framework makes the target question explicit by specifying five attributes: treatment conditions, target population, variable, intercurrent-event strategies, and population-level summary. The estimand is the precisely defined treatment-effect quantity of interest. In a Bayesian trial, a posterior distribution can express uncertainty about that quantity, but the posterior does not by itself define which quantity the trial was meant to estimate.
The attributes must work together. “Change in symptom score” names a variable but not its measurement time, target population, or comparison. “The average treatment effect” does not settle what happens if a participant stops assigned treatment or receives rescue medication. And a detailed analysis model still cannot resolve an undefined clinical question: it may estimate a quantity, but that quantity might not be the intended one.
A Compact Mathematical Representation
Let \(a\) and \(c\) denote the experimental and comparator treatment conditions. Let \(\mathcal P\) denote the target population, and let \(Y^{a,g}\) and \(Y^{c,g}\) denote outcomes under those conditions after applying the specified intercurrent-event strategy or strategies, summarized here by \(g\). This notation is shorthand: in practice, the strategy is specified for each relevant event and may define different outcome information for different events.
For a continuous endpoint, one possible population-level summary is the difference in means:
This equation does not define a complete estimand unless its symbols are given clinical meaning. The treatment conditions specify what \(a\) and \(c\) entail; the population specifies \(\mathcal P\); the variable and its time point specify \(Y\); the event strategy specifies \(g\); and the difference in means specifies the summary. For a binary endpoint, a risk difference, risk ratio, or odds ratio would instead define a different summary.
The estimand is a target quantity, not an observed data set or a calculation. An estimator is a rule for using data to estimate the target; an analysis model supplies assumptions and a likelihood for that analysis. Bayesian analysis combines a likelihood with a prior to produce a posterior for model parameters. For the posterior to address the clinical question, those parameters and the model must correspond to the specified estimand.
How the Attributes Fit Together
The following order is useful when writing a specification. It is not a substitute for clinical judgment: the details need to reflect the question stakeholders want answered.
Define the experimental and comparator regimens sufficiently to distinguish them, including relevant dose, schedule, duration, and treatment setting.
Specify the patients to whom the treatment comparison is intended to apply, rather than relying on an undefined phrase such as “trial patients.”
Name the outcome, measurement method, and assessment time or window. State any transformations or derived endpoint rules needed to make the outcome unambiguous.
Identify events after treatment initiation that affect either the existence or interpretation of the outcome, and state how each is handled in the target question.
State the contrast and scale, such as a mean difference or risk difference, and indicate the direction that represents benefit where appropriate.
An intercurrent event is an event occurring after treatment initiation that affects either the interpretation or existence of measurements associated with the clinical question. The strategy is part of the estimand, not a label for a statistical method used later to handle missing data. Common strategy names include treatment policy, hypothetical, composite, principal stratum, and while on treatment. Selecting among them requires a clinical question; a name alone does not complete the specification.
Worked Example: Continuous Symptom Score at a Fixed Time
Hypothetical scenario. A randomized trial compares a new oral medicine with placebo in adults meeting a stated symptom-severity and diagnostic eligibility definition. The intended endpoint is symptom-score change from baseline to week 12, with lower scores indicating fewer symptoms. For this example, the question is the mean difference in week-12 change regardless of whether participants discontinue assigned treatment or use rescue medication. The strategy for those events is therefore treatment policy.
Five attributes. The conditions are the specified oral medicine regimen and placebo, each added to the same background care. The target population is eligible adults as defined above. The variable is change in symptom score from baseline to week 12. The intercurrent-event strategy is treatment policy for discontinuation and rescue medication. The summary is experimental minus placebo mean change in that population.
Suppose, purely for illustration, the specified population-level mean changes are \(-4.2\) points under the medicine condition and \(-2.7\) points under placebo. Then the target contrast is
The negative difference indicates 1.5 points greater mean improvement in the experimental condition, given the stated score direction. These illustrative values specify and interpret the target contrast; they are not presented as results from an actual trial, and they do not establish how accurately a particular analysis would estimate it.
Justification: Why Changing an Attribute Changes the Target
The dependency on the attributes follows directly from the definition of the contrast. The mean \(E(Y^{a,g}\mid\mathcal P)\) depends on the treatment condition \(a\), the population distribution \(\mathcal P\), the endpoint \(Y\), and the strategy \(g\). The comparator mean has the same dependencies. Their difference therefore cannot be assumed to stay fixed when one of those inputs changes.
For example, changing from a week-12 outcome to a week-24 outcome replaces \(Y\) with a different measured variable. Restricting the population changes the distribution over which the means are taken. Changing the event strategy can change which outcomes the counterfactual quantities represent. Finally, replacing a mean difference with a ratio changes the scale of the summary. In each case the target is, in general, a different estimand. This is why an estimand should be fixed before debating which analysis model best estimates it.
Worked Example: A Binary Response and a Risk Difference
Hypothetical scenario. Suppose the target population is eligible adults with a specified chronic condition. The trial compares a 10-week investigational regimen with a specified control regimen. The variable is whether each participant achieves a pre-defined response by week 10, with response status evaluated regardless of treatment discontinuation. The chosen strategy for discontinuation is treatment policy. The population-level summary is the experimental minus control response probability.
For a numerical illustration, suppose the response probabilities defined by this target question are \(0.58\) and \(0.44\). The risk difference is
The contrast is 0.14, or 14 percentage points, higher response probability under the experimental condition for this population and endpoint. It is not a 14% relative increase: that would be a different summary. If the question instead called for a risk ratio, the corresponding calculation would be \(0.58/0.44=1.318\), about 1.32. The same two probabilities thus give different numerical summaries, each answering a different question.
Assumptions, Diagnostics, and Limits
Defining an estimand does not guarantee that trial data identify it. Interpretation requires a connection between the target question, trial conduct, measured data, and analysis assumptions. In a randomized comparison, randomization supports comparability of treatment groups at baseline in expectation, but does not fix outcome measurement, prevent post-randomization events, or ensure complete follow-up. A causal interpretation also relies on consistency (the observed outcome under a received condition agrees with the corresponding defined potential outcome), appropriate treatment assignment and follow-up, and sufficient information for the target comparison. The details needed to justify these conditions depend on the trial.
Useful checks include asking whether the population criteria are reproducible, whether treatment versions are defined, and whether the endpoint and time window can be applied consistently. Check that each relevant intercurrent event has an explicit strategy and that the proposed analysis does not silently substitute another strategy. For a Bayesian analysis, also check that the model parameters map to the specified population summary and that the data and model assumptions support the intended posterior interpretation.
Warning signs include several incompatible definitions of the primary endpoint in different documents; an analysis population that does not match the target population without a stated rationale; treating post-randomization exclusions as if they were baseline eligibility rules; or interpreting an analysis of observed outcomes after discontinuation as answering a hypothetical no-discontinuation question without additional assumptions. Missing data are not automatically an intercurrent-event strategy: how missing observations are handled in an analysis and what treatment question is targeted are related but distinct decisions.
A concise comparison helps keep the concepts separate:
| Item | What it specifies | Example |
|---|---|---|
| Estimand | The target treatment-effect quantity | Week-10 response risk difference in an eligible population under stated event strategies |
| Estimator | A data-based rule for estimating that quantity | A difference between estimated response probabilities |
| Analysis model | A probability model and assumptions used to analyze data | A model for response conditional on treatment and covariates |
| Decision rule | A rule for acting on analysis results | A posterior-probability threshold for declaring success |
The estimand is also not the same as the treatment effect observed among trial completers. That subset may differ from the target population because completion occurs after randomization. Nor does specifying five attributes resolve every practical issue: a target may be difficult to identify from available data, depend on strong assumptions, or have limited relevance to a particular decision. Those concerns should be made visible rather than hidden by precise notation.
Worked Example: Spotting a Mismatch Between Question and Analysis
Hypothetical scenario. A protocol asks for the mean difference in a week-16 functional score in all randomized eligible patients, regardless of treatment discontinuation. A report instead compares week-16 scores only among participants who remained on assigned treatment and had a recorded assessment. The report gives means of 31.0 and 34.5 points for experimental treatment and control, respectively.
The arithmetic difference is \(31.0-34.5=-3.5\) points. If lower scores mean better function, this is a 3.5-point lower mean in the analyzed completers. It is not automatically the estimand in the protocol. The restriction to observed completers and the treatment discontinuation pattern may change the population or event interpretation.
The discrepancy cannot be repaired by relabeling the reported contrast. The analysis would need a defensible connection to the stated target, with explicit assumptions and methods, or the target question would need to be revised transparently. The reported numbers alone do not identify the mean difference for all randomized eligible patients under the specified strategy.
Common Pitfall and Practical Review
A common mistake is to say, “The estimand is the endpoint analyzed with the Bayesian model.” That reasoning confuses the measurement, the analysis, and the clinical target. The correction is to write the treatment conditions, target population, outcome variable and time, event strategy, and population-level summary before choosing the analysis. Then verify that the model and data can support that target. A posterior probability about an ill-defined or mismatched parameter does not make the clinical question precise.
A useful final check is to read the specification as a complete sentence: “In [population], compare [treatment conditions] on [variable and time], handling [intercurrent events] by [strategies], and summarize the comparison as [population-level contrast].” If a reader could reasonably supply two different answers for any bracketed part, the estimand needs further definition.
Check Your Understanding
Answer each question, then check the short solutions below.
- A protocol names a mean difference in pain score but gives no assessment time. Which attribute is incomplete, and why does it matter?
- For a binary endpoint, hypothetical response risks are 0.62 under treatment and 0.50 under control. Calculate the risk difference and interpret it in percentage points.
- Explain why a posterior distribution is not itself an estimand.
- A study intends to describe all eligible patients regardless of treatment discontinuation, but analyzes only participants who complete treatment. What mismatch should be investigated?
- Name the five attributes required to specify the estimand.
Solutions. (1) The variable is incomplete because its assessment time is unspecified; response at different times can define different outcomes and treatment effects. (2) The risk difference is \(0.62-0.50=0.12\), or 12 percentage points higher response probability under treatment for the specified population and event strategy. (3) The estimand defines the target treatment-effect quantity; a posterior distribution represents uncertainty about model parameters or quantities linked to that target under a prior and likelihood. (4) Restricting to completers may change the population being compared and the interpretation of treatment discontinuation; the analysis must be justified as addressing the stated target or the target must be reconsidered. (5) Treatment conditions, target population, variable, intercurrent-event strategies, and population-level summary.