Estimands for Intercurrent Events in Pragmatic Clinical Trials

Pragmatic clinical trials ask what happens when an intervention is offered in routine care rather than under idealized conditions. Participants may not start treatment, may discontinue it, switch to another therapy, receive rescue medication, or encounter an outcome that changes the meaning of the planned endpoint. These events are not merely statistical inconveniences. They determine what treatment effect the trial result represents.

The ICH E9(R1) estimand framework provides a way to align the clinical question, target population, treatment conditions, outcome variable, intercurrent-event strategy, and population-level summary. In a pragmatic trial, this alignment is especially important because treatment adherence and co-interventions are part of the setting the trial intends to study. Removing those events from the primary analysis can change an effectiveness question into an efficacy question.

1. Define the pragmatic clinical question

Begin with the decision the trial is intended to inform. A health system may ask whether offering a new care pathway improves outcomes under routine uptake. A clinician may ask whether prescribing a therapy produces benefit despite imperfect adherence and rescue treatment. A regulator may ask about the effect of assigning a product under specified conditions. These questions can involve the same intervention but require different estimands.

Describe the treatment conditions as interventions, not only as labels such as “drug A” and “drug B.” Include initiation, allowed titration, monitoring, treatment duration, co-interventions, and access to rescue therapy. If the intervention is a care pathway, specify which components are delivered by the system and which remain subject to clinician or patient choice.

The target population should represent the people and care setting for the decision. State eligibility, recruitment frame, baseline time, site context, and whether the trial population is intended to represent routine service users or a narrower clinical group. The estimand cannot be more general than the population and setting supported by the design.

2. Distinguish intercurrent events from missing outcomes

An intercurrent event occurs after treatment initiation or assignment and affects either the interpretation or the existence of the outcome variable for the clinical question. Examples include discontinuation, treatment switching, use of rescue therapy, pregnancy, death, and initiation of a competing intervention. A missing outcome is an unobserved measurement. The two problems can be related but should not be conflated.

If a participant stops treatment but continues follow-up with outcome measurement, the outcome is not automatically missing. A treatment-policy estimand may retain that participant’s data because the question concerns assignment in routine practice. If a participant dies before a later quality-of-life measurement, the measurement may not exist in the same form, and a strategy such as a composite or a carefully defined hypothetical estimand may be needed.

The statistical analysis plan should state what data are collected after each intercurrent event. A strategy cannot be implemented reliably if the event is not recorded, the timing is unknown, or outcome follow-up stops when treatment changes even though the primary estimand requires continued observation.

3. Treatment-policy strategy for real-world effectiveness

The treatment-policy strategy estimates the effect of assigning participants to treatment conditions regardless of subsequent adherence, switching, rescue therapy, or discontinuation, provided outcomes continue to be measured according to the protocol. It often aligns with an intention-to-treat interpretation and is a natural primary estimand for pragmatic effectiveness questions.

The treatment-policy effect answers a policy question: what is the difference in outcomes between offering or assigning these strategies in the target setting, with the adherence and co-intervention patterns that follow? It does not estimate the effect that would occur if every participant perfectly followed treatment.

To support this estimand, continue outcome ascertainment after treatment changes and record deviations accurately. An analysis based only on participants who remain on assigned treatment does not estimate treatment policy unless the estimand and assumptions explicitly justify that restriction. Per-protocol analyses may be supplementary estimands, but they should not be presented as interchangeable with treatment-policy results.

4. Hypothetical strategy for a controlled scenario

The hypothetical strategy asks what the outcome would have been if an intercurrent event had not occurred or had occurred differently. For example, the trial may estimate the effect if treatment discontinuation had not occurred, or if rescue medication had not been initiated. This can be clinically useful when the question concerns pharmacologic efficacy under sustained exposure.

Hypothetical estimands require assumptions about outcomes that are not observed after the event. The analysis may use models, imputation, censoring adjustments, or other methods to estimate that counterfactual trajectory. Sensitivity analyses should vary the assumptions because the data after discontinuation cannot directly reveal what would have happened without it.

In pragmatic trials, a hypothetical primary estimand may be poorly aligned with the purpose of studying routine care. It can still be valuable as a supplementary estimand when decision-makers need to understand the effect under a specified controlled scenario. The protocol should explain why that scenario matters and how it differs from the treatment-policy question.

5. Composite strategy when the event is clinically part of the outcome

A composite strategy incorporates an intercurrent event into the outcome variable. Death, hospitalization, treatment failure, rescue therapy, or permanent discontinuation may be combined with a later clinical outcome when the combined result has a coherent clinical interpretation.

Composite outcomes require careful ordering and interpretation. A participant who dies cannot later contribute a symptom score in the same way as a survivor, so a composite may avoid assigning an undefined outcome. However, components can differ in importance, frequency, and treatment effect. Report component-specific results and explain whether the composite is driven by a clinically meaningful component or by a frequent but less important event.

Do not use a composite simply to avoid missing-data decisions. The components should represent related clinical consequences of the treatment strategy. Define the time window, event hierarchy, and handling of multiple events before the trial begins.

6. Principal-stratum strategy for a defined subgroup

The principal-stratum strategy defines the treatment effect within a subgroup characterized by post-randomization potential outcomes, such as participants who would adhere under either treatment condition or who would not require rescue therapy under either strategy. This can address a scientifically interesting question, but the relevant subgroup is usually not directly observable because each participant experiences only one randomized condition.

Principal-stratum estimands therefore require strong assumptions, additional modeling, or instrumental-variable-like reasoning. They should not be presented as ordinary per-protocol analyses. If the subgroup cannot be identified credibly, report the uncertainty and consider whether a treatment-policy estimand better matches the clinical decision.

The strategy can be useful when the decision explicitly concerns a latent group, but the protocol must explain why that group matters. The analysis should distinguish observed adherence from the potential adherence behavior that defines the principal stratum.

7. While-on-treatment strategy and its limits

The while-on-treatment strategy focuses on outcomes while participants remain on the assigned treatment, often censoring after discontinuation or switching. It can answer a treatment-exposure question, but informative censoring is a central concern because health status and early response may influence the decision to stop or change treatment.

If the analysis censors participants at treatment deviation, the remaining participants may no longer be comparable across arms. Inverse probability of censoring weighting, g-methods, or other methods may be needed under appropriate exchangeability and positivity assumptions. The weights must be diagnosed, and the assumptions should be stated rather than hidden inside a software procedure.

While-on-treatment results are often interpreted as if they were treatment-policy effects. That interpretation is incorrect unless the treatment-policy and while-on-treatment questions coincide under the study design. Report both when useful, but label their clinical meanings clearly.

8. Match estimand strategy to pragmatic trial operations

Estimand choice should influence data collection. If the primary question uses treatment policy, capture outcomes after switching, discontinuation, rescue therapy, and non-adherence. If the analysis uses a composite strategy, collect precise dates and component definitions. If a hypothetical strategy is important, record the event that triggers the counterfactual analysis and the covariates needed for sensitivity analysis.

Site workflows should use a common event taxonomy. Treatment discontinuation because of an adverse event differs from discontinuation because of remission, access, preference, or administrative error. The same event label can have different causal implications if its reason and timing are not documented.

Protocol deviations should be visible to clinicians, statisticians, data managers, and the patient advisory group. The estimand table can serve as a shared map from the clinical question to the event definition, follow-up procedure, analysis set, and interpretation of the final estimate.

9. Evidence summary table

10. Actionable Steps: Build an Estimand-to-Analysis Plan

11. Sensitivity analysis is not a replacement estimand

A sensitivity analysis examines how robust an estimate is to assumptions used to estimate the same estimand. A supplementary analysis estimates a different estimand. These activities should be labeled separately. For example, a treatment-policy primary estimand may have a sensitivity analysis under alternative missing-data assumptions, while a hypothetical efficacy estimand may be a separate supplementary question.

Use sensitivity analysis to address assumptions that cannot be verified from observed data. For a hypothetical strategy, vary the post-event outcome distribution. For while-on-treatment analyses, vary the censoring model and the strength of unmeasured predictors of deviation. For a composite, examine whether the conclusion changes when components are analyzed separately or ordered differently.

Interpretation should identify which assumptions change the decision. A stable point estimate with a wide range of clinically plausible sensitivity results is not robust merely because the primary p-value is below a threshold. Report the estimand, assumptions, analysis, and conclusion together.

12. Common interpretation errors

A common error is calling every analysis “intention-to-treat” without stating which intercurrent events are included and how they are handled. Randomization supports a treatment-policy effect only when follow-up and analysis preserve the assigned-strategy question. It does not automatically answer a hypothetical or while-on-treatment question.

Another error is treating treatment discontinuation as missing data by default. A participant can stop treatment and still provide an observed outcome that is relevant to treatment policy. Conversely, death can make a later outcome undefined, which may require a composite or another explicitly defined strategy.

Do not use an estimand label without an operational definition. “Per protocol,” “on treatment,” and “as treated” can mean different things across studies. Specify the event, time, censoring rule, analysis population, and assumptions so readers can reproduce the target question.

Conclusion

Estimands make pragmatic clinical trials interpretable by stating what treatment effect is being estimated when real-world events occur after assignment. Treatment policy often provides the most direct effectiveness question, while hypothetical, composite, principal-stratum, and while-on-treatment strategies answer different clinical questions with different assumptions. A credible trial defines events before data collection, aligns follow-up and analysis with each strategy, distinguishes sensitivity analyses from supplementary estimands, and reports the boundaries of interpretation. The estimand-to-analysis table is a practical tool for ensuring that the clinical objective survives contact with real-world adherence, switching, rescue therapy, discontinuation, and competing outcomes.