A modeling discussion can become highly technical before the team has agreed on the treatment effect it wants to estimate. That sequence creates avoidable ambiguity. A precise estimate is not automatically an estimate of the intended clinical question.
ICH E9(R1) establishes the estimand framework. The working exercise below is our recommendation for carrying that framework into a practical review meeting, rather than turning it into another disconnected document.
1. Write the clinical question in plain language
Consider a hypothetical randomized study in which some participants begin rescue medication. Is the objective to estimate the effect of assignment under the treatment conditions that occur, or an effect under a scenario in which rescue was unavailable? Those are different questions, even if they use the same endpoint variable.
State the treatment conditions, target population, variable, handling of relevant intercurrent events, and population-level summary. Discuss the question with clinical, statistical, and operational colleagues before committing to an implementation.
2. Distinguish events from absent observations
Treatment discontinuation is an event that may affect interpretation. A missing assessment is a lack of information. The two can occur together, but they should not be treated as synonyms. If data collection stops after treatment discontinuation, the collection policy may make the intended treatment-effect question harder to answer.
Review the assessment schedule and follow-up plan alongside the estimand. Ask which observations are needed under each strategy and whether sites can feasibly collect them. This links a statistical choice to trial operations rather than leaving the consequences until analysis.
3. Choose an estimator with defensible assumptions
Once the question is clear, evaluate methods that target it and identify their assumptions. Explain how observed data enter the estimator, what must be assumed about unobserved outcomes, and how uncertainty will be represented. A familiar model is not a sufficient reason to select it.
The EMA missing-data guideline is useful additional reading. Our recommendation is to write an assumption register with the clinical rationale, supporting information, and planned challenges to each material assumption.
4. Keep sensitivity analyses aligned
A sensitivity analysis should probe assumptions relevant to the same estimand. An analysis that addresses a different treatment-effect question may instead be supplementary. Label the distinction so that a set of alternative results is not mistakenly presented as one interchangeable collection.
ICH E8(R1) supplies a broader study-design context. For a practical handoff, attach the approved question, intercurrent-event decisions, estimator, and sensitivity strategy to the corresponding analysis specification.
The value of this work becomes apparent during interpretation. When the team understands what was estimated, it can discuss the result’s clinical relevance and limitations without reconstructing the statistical intent from code or table footnotes.
