Estimands and Intercurrent Events in Clinical Trials: Defining the Treatment Effect Precisely

A clinical trial question often sounds simple: does the investigational treatment improve the primary outcome compared with control? The question becomes less simple when participants discontinue treatment, receive rescue medication, switch therapies, die before the planned endpoint, or deviate from the assigned regimen. These events are not merely missing observations. They can change the meaning of the treatment effect the trial is trying to estimate.

The ICH E9(R1) addendum on estimands and sensitivity analysis introduced a structured way to resolve this ambiguity. Instead of selecting an analysis first and explaining its interpretation later, investigators define the treatment-effect question using five attributes. The estimand then determines which data are relevant, how intercurrent events are handled, and which sensitivity analyses are needed to support the conclusion.

1. The five attributes of an estimand

A complete estimand describes the treatment effect for a defined population, under specified treatment conditions, on a defined variable or endpoint, while explicitly addressing intercurrent events and stating the population-level summary measure. The population might be all randomized participants or an eligible subgroup. Treatment conditions may describe treatment assignment, treatment received, or a treatment strategy that includes rescue medication.

The endpoint must be defined with enough precision to identify how it is measured and when it is assessed. The intercurrent-event strategy then explains what the outcome means if a participant discontinues treatment, uses rescue therapy, or experiences another event that affects interpretation. Finally, the summary measure may be a mean difference, risk ratio, odds ratio, hazard ratio, or another clinically interpretable contrast.

2. Why intercurrent events change the question

Suppose a patient assigned to an active treatment stops because of intolerance and then receives rescue therapy. An analysis that ignores post-discontinuation outcomes asks a different question from one that follows the patient regardless of treatment changes. Neither question is automatically correct. The appropriate choice depends on whether the clinical decision concerns starting the treatment, adhering to it, using it as part of a treatment strategy, or understanding the biological effect while treatment is received.

The estimand framework makes this distinction explicit. It prevents a common reporting error in which the protocol describes one treatment effect, the analysis estimates another, and the conclusion uses a third interpretation. By writing the estimand before database lock and unblinding, the study team can align protocol objectives, data collection, statistical analysis, and clinical interpretation.

3. Evidence summary table

4. The five strategies for intercurrent events

• Treatment policy: Include the outcome regardless of whether the intercurrent event occurs. This strategy answers the effect of assigning a treatment in real-world use, including discontinuation or rescue therapy.
• Composite strategy: Incorporate the intercurrent event into the endpoint. For example, death or a treatment failure before the primary assessment may be included as part of a composite outcome.
• Hypothetical strategy: Estimate what would have happened if the intercurrent event had not occurred. This requires assumptions or models for outcomes under an unobserved scenario.
• While-on-treatment strategy: Define the effect up to the time the treatment is discontinued or rescue therapy begins. This may describe biological treatment response but can be sensitive to informative discontinuation.
• Principal-stratum strategy: Estimate the effect within a subgroup defined by potential responses to the intercurrent event under each treatment. This can be clinically meaningful but is often difficult to identify without strong assumptions.

5. Estimands, estimators, and analyses are different

An estimand is the target question; an estimator is the statistical procedure used to estimate it; and an analysis is the complete implementation, including data handling, model assumptions, and sensitivity analyses. A mixed model, multiple imputation procedure, or inverse-probability weighting method does not define the estimand by itself. The same estimator can support different estimands depending on how intercurrent events and outcomes are defined.

This distinction is especially important for treatment-policy questions. An intention-to-treat analysis is often aligned with a treatment-policy estimand, but the label alone is not enough. Investigators must specify how rescue therapy, treatment switching, death, and post-discontinuation measurements enter the endpoint and model. Conversely, a per-protocol analysis may target a hypothetical or while-on-treatment question, but only if the intervention strategy and deviations are defined precisely.

6. Actionable steps for an estimand-driven trial

7. Sensitivity analysis is part of the evidence claim

Some estimands require assumptions about outcomes that are not observed after an intercurrent event. A hypothetical strategy may require modeling outcomes under a counterfactual treatment path. A while-on-treatment strategy may require assumptions about informative discontinuation. Sensitivity analyses should vary these assumptions in clinically plausible directions and show whether the conclusion is stable.

The goal is not to find a single analysis that produces a preferred result. The goal is to describe how robust the treatment-effect conclusion is to assumptions that cannot be verified directly. A result that changes materially across plausible sensitivity scenarios should be reported with appropriate uncertainty rather than presented as a definitive clinical benefit.

8. Common errors in manuscripts

Frequent problems include treating treatment discontinuation as ordinary missing data without explaining the target question, using an estimand label without defining the intercurrent-event strategy, and presenting a per-protocol result without describing how treatment deviations were modeled. Another error is applying a composite endpoint simply because it avoids missing outcomes, without explaining whether the components have comparable clinical importance or whether the composite answers the intended decision question.

A strong manuscript presents an estimand table before the primary analysis results. The table should identify the population, treatment conditions, endpoint, intercurrent-event strategy, summary measure, estimator, and key sensitivity analyses. This structure allows clinical readers, statisticians, regulators, and peer reviewers to determine whether the analysis actually answers the question the trial set out to address.

Conclusion

Estimands turn vague treatment-effect language into a precise clinical question. By identifying intercurrent events before analysis and choosing an explicit strategy for each one, investigators can align trial objectives, data collection, statistical methods, and interpretation. ICH E9(R1) does not prescribe one universal estimand; it provides the discipline needed to choose an estimand that matches the clinical decision. Trials that make this choice transparently are easier to analyze, easier to interpret, and more credible to patients, regulators, and peer reviewers.