;)
Estimands in Clinical Trial Design: A Framework That Changes How We Answer the Research Question
A clinical trial answers a question. But what question, precisely? The ICH E9(R1) addendum on estimands—adopted in November 2019 and now widely implemented—has revealed how often this supposedly simple question remains fundamentally ambiguous until explicitly formalized (Stanworth et al. 2024).
An estimand is the precise parameter a trial intends to estimate. It comprises five attributes: treatment, population, endpoint, intercurrent event strategy, and population-level summary.
While this framework seems technical, it solves a substantive problem. Intercurrent events are events occurring after randomization that affect the endpoint, such as treatment discontinuation, concomitant medication use, and disease progression despite treatment. These events can be handled in multiple ways, each yielding different treatment effect estimates (Wason et al., 2020).
The practical consequence: without explicit estimand definition, the same trial data analyzed with different (but reasonable) intercurrent event strategies will produce contradictory answers to what appeared to be the same research question (BMJ, 2024).
Five strategies now exist: treatment policy (including all post-randomization observations), hypothetical (estimating treatment effect if the intercurrent event wouldn’t occur), composite (treating the event as part of the outcome), while on treatment (restricting observation to before the event), and principal stratum (defining population-specific effects) (Wason et al., 2020).
These five strategies are explored further in Table 1, below.
Table 1: Intercurrent Event Strategies Used in The Estimand Framework
| Strategy | Question it answers (informally) | How it handles the intercurrent event | Typical use case |
| Treatment policy | “What is the effect of assigning treatment in usual practice, including nonadherence?” | Keeps all post event data in the analysis | Pragmatic/real world effect for regulators, HTA bodies, payers |
| Hypothetical | “What would the effect be if the event did not happen?” | Conceptually removes the event; often uses modelling or imputation | Understanding pure biological effect; scenarios not feasible in practice |
| Composite | “What is the effect when the event itself counts as part of the outcome?” | Recodes the event into the endpoint (e.g. event = failure) | When the event is clinically equivalent to a bad outcome (death, relapse, etc.) |
| While on treatment | “What is the effect while patients are actually receiving treatment?” | Uses only data up to the time of the event | Safety/efficacy during exposure; early benefit/risk evaluation |
| Principal stratum | “What is the effect in the subgroup defined by whether the event would (or would not) occur?” | Restricts to a conceptual subpopulation defined by event potential | Causal/ mechanistic questions; when event strongly tied to biology or prognosis |
Consider a Phase III type 2 diabetes trial. Patients discontinue treatment. Standard intention-to-treat analysis—following all randomized patients to endpoint regardless of adherence—captures a “real-world” treatment effect including discontinuation. A hypothetical estimand, by contrast, estimates the effect if patients remained on treatment despite discontinuation. An alternative composite estimand could classify discontinuees as treatment failures. Each answers a different research question; all are scientifically valid; all are practically relevant to different audiences (regulators, health systems, clinical decision-makers) (Stanworth et al., 2024).
Regulatory agencies now demand explicit estimands. EMA scientific advice requests increasingly scrutinize whether sponsors have clearly defined their primary estimand and justified their intercurrent event strategy (Mészáros et al., 2024). FDA and international regulatory bodies emphasize that estimands should be pre-specified, thoroughly justified, and directly connected to trial objectives, recognizing that rigorous estimand definition prevents post-hoc analysis and ensures alignment between study design and research questions.
Implementation of regulatory guidance on estimands requires a cultural shift. Specifying an estimand before trial execution forces uncomfortable precision. What exactly do we want to know? Whose perspective matters—regulators, patients, clinicians, health systems? What intercurrent events are anticipated? How should they be handled? These aren’t statistical questions; they’re clinical and strategic questions that statisticians can’t answer alone (Mészáros et al., 2024).
The estimands framework is not optional sophistication. It’s becoming mandatory intellectual honesty about what research questions a trial will actually address. Organizations that treat estimands as a regulatory checkbox rather than a fundamental design element put themselves at risk of producing trials whose analyses are defensible, but whose results may be ambiguous (Stanworth et al., 2024).
References
Mészáros, L., Lasch, F., Delafont, B., & Guizzaro, L. (2024). Estimands in CNS trials—A review of strategies for addressing intercurrent events. Contemporary Clinical Trials Communications, 38, 101266. https://doi.org/10.1016/j.conctc.2024.101266
Stanworth, S. J., Wells, A. W., Contreras, M., & Oakley, F. (2024). The estimands framework: A primer on the ICH E9(R1) addendum. British Medical Journal, 384, bmj-2023-076316. https://doi.org/10.1136/bmj-2023-076316
Wason, J. M. S., Steinsaltz, D., & Flack, N. P. (2020). A narrative review of estimands in drug development and regulatory evaluation of therapeutics. Statistics in Biopharmaceutical Research, 12(2), 169–177. https://doi.org/10.1080/19466315.2020.1744090
