The principal stratum strategy estimates a treatment effect in a population defined by the potential occurrence or non-occurrence of an intercurrent event under the treatments being compared. It is not a subgroup analysis based on baseline characteristics; rather, it focuses on a stratum described by possible post-randomization trajectories. This allows it to address a clinically narrowly defined question, but it is methodologically demanding because it relies on unobservable counterfactuals.
Population defined by potential trajectories
A principal stratum may, for example, include individuals who would not require a specific rescue medication under either of the treatments being compared. However, for each individual, only the trajectory under the treatment actually assigned is observable. Whether the same person would have needed rescue medication under the other treatment remains unobserved; it is precisely this missing information that distinguishes the strategy from an analysis of the observed event groups.
The estimand must specify which event defines the stratum, which potential event trajectories are included, and which variable the comparison refers to. A wording such as “effect in people without treatment discontinuation” is not sufficient if it merely refers to those who did not discontinue in the dataset of their respective arm. That would be selection based on an observation that arises after randomization.
Identifiability and assumptions
Randomization does not automatically identify the effect within a principal stratum because stratum membership under both treatments is not fully observed. Estimation therefore requires additional assumptions, modelling, or information. The scientific rationale must therefore explain why the assumed relationship between the potential event trajectories and the endpoint is plausible for the specific indication.
The strategy can help avoid bias caused by a post-randomization event when the target population is truly intended to be defined by potential trajectories. However, it does not create an observable subpopulation by simply filtering the dataset. ICH E9(R1) explicitly notes that practical feasibility and the strength of the required assumptions must be assessed before choosing this strategy.
The choice of event definition determines whether the stratum remains clinically interpretable. A very broad category such as “any treatment consequence” bundles causes with different meanings; a narrow event, by contrast, may result in too few individuals in the relevant stratum. Time windows and decision rules must therefore specify whether, for example, rescue medication before the primary visit, permanent discontinuation, or a specific event profile is decisive. Without this specification, the potential trajectories are not described unambiguously.
Accordingly, the interpretation concerns a target population that is not fully observable. It must not be presented as an ordinary mean among participants who were actually event-free.
Distinction from subgroup analysis
Unlike subgroup analysis, the principal stratum strategy does not use characteristics collected before randomization, such as age, severity, or biomarker status. Such baseline variables allow the randomized individuals to be classified before the treatment course unfolds. A principal stratum, by contrast, is based on how an intercurrent event would behave under one or both hypothetical treatments.
The established concept of subgroup analysis therefore addresses a different scientific question: the possible heterogeneity of an effect in groups that can be described in advance. A retrospectively observed group of individuals who did not receive rescue medication in their arm is neither a principal stratum nor a standard baseline subgroup. Directly comparing such groups can undermine the comparability created by randomization.
Relevance for clinical trials
Before the study starts, the sponsor and statistics team must clarify whether the intended subpopulation is clinically decision-relevant and whether the event definition, time window, and endpoint collection can support the required assumptions at all. For reporting results, a point estimate is not sufficient: dependence on the set of assumptions and the limits of identifiability are part of the interpretation, especially when the event rate is high.
Full-service CROs such as Mediconomics support the precise definition of the event and the target population, plan the capture of the relevant time points and concomitant therapies, document the model assumptions in the statistical analysis plan, and coordinate tables and explanatory text for the clinical study report.
Frequently Asked Questions (FAQ)
Can a principal stratum be directly flagged in the dataset?
Generally, no. Only whether the event occurred under the assigned treatment is observable. Membership in a stratum defined across both potential trajectories cannot be fully observed.
Why is excluding all individuals with treatment discontinuation not a principal stratum analysis?
Such exclusion uses the event actually observed after randomization. A principal stratum question, by contrast, requires a definition based on the potential event trajectories under the treatments being compared.
When is this strategy particularly difficult to justify?
It is critical when the event is strongly influenced by treatment and there is little substantive basis for assumptions about the unobserved trajectory under the other treatment.
Regulatory References
- ICH E9(R1), Addendum on Estimands and Sensitivity Analysis in Clinical Trials – describes the principal stratum strategy and related interpretation issues.
- ICH E9, Statistical Principles for Clinical Trials – provides the framework for prospective statistical objectives.
- ICH E17, General Principles for Planning and Design of Multi-Regional Clinical Trials – addresses planning clearly defined target populations across regions.
