The hypothetical strategy is a way of handling an intercurrent event in the estimand as defined in ICH E9(R1). It describes the treatment effect in an explicitly formulated scenario in which that event would not have occurred—for example, if rescue medication had not been available. The estimated value is therefore not directly observable in all randomised participants, but is based on assumptions about the unobserved course.
Formulate the hypothetical question precisely
The strategy requires more than the statement “without treatment discontinuation”. The protocol must specify which event is assumed to be prevented, to whom the scenario applies, and which endpoint is compared under which treatment conditions. For a diabetes endpoint, for example, the question may concern the difference in the HbA1c trajectory if no rescue medication had been initiated; in the case of a treatment switch, the clinical relevance of foregoing subsequent treatment must be justified separately.
What is crucial is the distinction between an endpoint that remains conceivable and a variable that does not exist. After death, a later functional score cannot be transferred into a scenario with merely no treatment discontinuation, because the measurement cannot be taken. ICH E9(R1) therefore notes that a hypothetical estimand cannot be constructed for variables that are not measurable due to death.
Estimation beyond what is observed
After the intercurrent event occurs, observed values are not simply discarded. Depending on the research question, models for repeated measures, multiple imputation, inverse weighting, the g-formula, or g-estimation can be used to estimate the trajectory in the defined scenario. Which method fits depends on which post-randomisation covariates are intended to explain the probability of the event and the endpoint trajectory.
Every method requires assumptions that cannot be fully verified from the study data. If, for example, participants with a less favourable course receive rescue medication more often, the model must capture the relevant precursors; otherwise, a missing value will be replaced with an overly favourable trajectory. Sensitivity analyses deliberately vary such assumptions and assess how strongly the conclusion depends on plausible deviations.
The analysis may also use values after the event, provided their relationship to the intended scenario and to the chosen model is set out. Blanket truncation of all measurements after an add-on therapy does not automatically make the estimation hypothetical; it initially only changes the information available and can increase bias from prognostically driven events.
Distinction: not a treatment-policy strategy
Unlike the treatment-policy strategy, the hypothetical strategy does not estimate an effect that includes the intercurrent event as it actually occurred. The treatment-policy strategy uses the observed endpoint values regardless of whether, for example, add-on medication or a treatment switch has occurred. It therefore preserves a different, more pragmatic clinical question than a scenario from which precisely that event is conceptually removed.
The hypothetical strategy is also not equivalent to a per-protocol analysis. A per-protocol analysis defines a population or an aspect of adherence; the hypothetical strategy, by contrast, specifies the counterfactual scenario for a specific intercurrent event within an estimand. Both can work with post-randomisation information, but they do not automatically answer the same question and must not be replaced by the same exclusion rule.
Relevance for clinical trials
For study planning, the strategy means that predictors of the intercurrent event must be collected in good time and with sufficient temporal resolution. For rescue medication, indication, start date, dose, and the clinical values immediately beforehand must be captured in the data; for treatment discontinuation, the reason and timing are relevant. In the SAP, the primary analysis, model assumptions, handling of post-event data, and sensitivity analyses must correspond to the specific hypothetical scenario.
The clinical wording must not implicitly assume that an event could be prevented without risk. A hypothetical question without rescue medication, for example, differs from the therapeutic decision to actually withhold that medication from a patient in the event of deterioration.
Full-service CROs such as Mediconomics support hypothetical strategies through estimand workshops with clinical and statistical functions, derivation of targeted CRF variables, planning of imputation or weighting analyses, and documentation of sensitivity assumptions. This makes it possible to trace in the protocol and CSR which effect is intended to be described without which event, and which data support that estimate.
Frequently Asked Questions (FAQ)
Is a hypothetical effect a “pure” drug effect?
Not necessarily. It only describes the effect under a precisely delimited scenario, for example without rescue medication. Other concomitant treatments, adherence issues, or events may remain part of the estimand if they are not also explicitly handled hypothetically.
Why are sensitivity analyses particularly important here?
The endpoint values after the event assumed to be prevented are not fully observed. Different plausible assumptions about these values or about unmeasured influencing factors can therefore yield different estimates; their impact must be reported separately.
Can the same study include a hypothetical estimand and a treatment-policy estimand?
Yes, if both clinical questions are relevant and each is clearly pre-specified. The endpoint variables, data collection, and analyses must then make clear which results belong to which handling of the respective intercurrent event.
Regulatory References
- ICH E9(R1), Addendum on Estimands and Sensitivity Analysis – defines the hypothetical strategy and its limitations in the event of death.
- EMA/CHMP/ICH/436221/2017, ICH E9(R1) Step 5 – provides the regulatory version applicable in Europe.
- ICH E6(R3), Good Clinical Practice – provides the rationale for the planned collection of data that support the analysis.