Estimand Framework

The estimand framework is a structured way of stating, in a clinical trial protocol, exactly what treatment effect the trial is designed to estimate — before the trial is analyzed, and ideally before it enrolls. It was introduced through the ICH E9(R1) addendum to the International Council for Harmonisation’s foundational statistical guideline, ICH E9, “Statistical Principles for Clinical Trials.” The addendum reached its final version in November 2019, following a 2017 draft and multi-year consultation, and is now the reference framework cited across biostatistics, regulatory submissions, and trial reporting for this issue.

The problem the framework addresses is that a single trial can support several different, equally legitimate treatment-effect questions, and without stating which one is intended, a reported result can be interpreted more precisely than the underlying analysis actually supports. This becomes concrete around missing data and treatment discontinuation: if a meaningful share of participants stop their assigned treatment, switch to a rescue medication, or are lost to follow-up, the ‘effect’ reported can mean substantially different things depending on how those post-randomization events are handled — and two trials, or two analyses of the same trial, that appear to disagree may in fact simply be answering different questions.

The five estimand attributes

ICH E9(R1) specifies that a fully defined estimand requires five attributes, agreed together as a set:

• Population — the patients targeted by the clinical question, ordinarily aligned with the trial’s inclusion/exclusion criteria.
• Treatment — the treatment condition of interest and the comparator it is being compared against, including relevant details of the treatment regimen (e.g. concomitant or subsequent therapy where relevant to the question).
• Variable (endpoint) — the specific outcome measured for each participant that addresses the clinical question, and, where relevant, how it is measured and at what timepoint.
• Handling of intercurrent events — the pre-specified strategy for events occurring after treatment initiation that affect either the interpretation or the existence of the variable’s measurements — most commonly treatment discontinuation, use of rescue or alternative therapy, treatment switching, or death.
• Population-level summary — the measure used to compare treatment conditions across the population, such as a difference in means, a risk difference, an odds ratio, or a hazard ratio.

Changing any one attribute — most consequentially, the intercurrent-event strategy — changes what the resulting effect estimate means, even when the same raw data and the same statistical model are used to compute it. This is the core discipline ICH E9(R1) is aiming for: the estimand comes first, defined from the trial’s clinical objective, and the statistical analysis method is chosen afterward specifically because it estimates that estimand — not the reverse.

Strategies for intercurrent events

ICH E9(R1) sets out several standard strategies for handling an intercurrent event within an estimand definition, each corresponding to a different underlying question:

• Treatment policy strategy — the value of the variable is used regardless of whether the intercurrent event occurred, and outcomes are attributed to the initially assigned treatment strategy. This answers a question close to “what is the effect of being assigned this treatment,” broadly consistent with an intention-to-treat orientation; see Intent-to-Treat (ITT) vs. Per-Protocol Analysis for how that related but distinct concept is defined.
• Hypothetical strategy — asks what the variable’s value would have been had the intercurrent event not occurred, treating the actual post-event data as unobserved and requiring it to be estimated, commonly through a statistical model.
• Composite variable strategy — the intercurrent event itself is incorporated into the definition of the variable (for example, treating treatment discontinuation as a component of a composite “treatment failure” outcome alongside the original clinical endpoint).
• While on treatment (while-alive) strategy — uses only the measurements taken up to the occurrence of the intercurrent event, addressing the treatment effect during the period participants remained on treatment.
• Principal stratum strategy — restricts the question to the subpopulation of participants who would not have experienced the intercurrent event under a specified treatment condition, a strategy that typically requires stronger assumptions to estimate than the others.

A protocol can, and often does, define a primary estimand using one strategy and one or more supplementary or sensitivity estimands using a different strategy for the same intercurrent event, to show how much the trial’s headline conclusion depends on that specific handling choice.

Why the framework matters for trial design and interpretation

Before ICH E9(R1), a protocol’s statistical analysis plan would typically specify an analysis population (such as intent-to-treat or per-protocol) and an analysis method, but not necessarily connect those choices back to a single, explicitly stated clinical question. The estimand framework requires that connection to be made explicit and made first: sponsors, regulators, and investigators agree on the clinical question an endpoint is meant to answer, then select the population, variable, intercurrent-event handling, and summary measure that together answer it, and only then choose an estimation method. This has direct consequences for protocol development, statistical analysis plans, and how missing data is handled — see Designing a Clinical Trial: Endpoints, Sample Size, Randomization, SAP for how estimand definition fits alongside endpoint selection and the broader statistical analysis plan, and Randomized Controlled Trial (RCT) for the underlying trial design the estimand framework applies to.

Frequently asked questions

Is an estimand the same as an endpoint?

No. The variable (or endpoint) is only one of the five estimand attributes. An estimand also specifies the population, the treatment conditions compared, how intercurrent events like treatment discontinuation are handled, and the population-level summary measure — the same endpoint can support different estimands depending on those other four attributes.

Why was the estimand framework introduced?

Regulators and statisticians observed that trials could report a treatment effect without clearly specifying which of several plausible clinical questions it answered, particularly regarding how discontinuation, treatment switching, or rescue medication use were handled. ICH E9(R1), finalized in November 2019, addresses this by requiring the estimand to be defined explicitly, aligned with the trial objective, before the analysis method is chosen.

How does the estimand framework relate to intention-to-treat analysis?

Intention-to-treat is an analysis principle, not itself a complete estimand. A treatment-policy strategy for intercurrent events, applied to the full analysis set, produces an estimand broadly consistent with the traditional intention-to-treat approach — but the estimand framework makes the underlying question and the handling of specific intercurrent events explicit in a way that the intention-to-treat label alone does not.

Does every trial need multiple estimands?

A trial typically defines one primary estimand corresponding to its primary clinical question, but frequently defines one or more supplementary or sensitivity estimands for the same or secondary endpoints, often varying the intercurrent-event strategy, to test how robust the primary conclusion is to that specific analytic choice.