The Estimand Framework: Defining Exactly What a Trial Is Trying to Estimate

An estimand is a precise, written-down definition of exactly what treatment effect a clinical trial is trying to estimate, specified before anyone looks at the data. The ICH E9(R1) addendum, adopted internationally in 2019 and made legally effective in the European Union on 30 July 2020, breaks that definition into five components and forces trialists to state, in advance, how they will handle the messy events that inevitably occur during a study. That single choice quietly determines what the reported number actually means, and two trials of the same drug can produce different effects simply because they defined different estimands.

The gap the framework was built to close

For decades, protocols described their analysis in terms of methods: intention-to-treat, per-protocol, last-observation-carried-forward. Those are recipes for handling data, but they do not clearly state the question. When a patient stops the study drug because of a side effect, starts a rescue medication, or dies before the final measurement, the "treatment effect" becomes ambiguous. Are we estimating the effect of the drug as prescribed regardless of what happens next, or the effect it would have had if nobody had stopped or switched? These are genuinely different questions with different answers, and the older vocabulary let people blur them.

ICH E9(R1), the addendum to the original 1998 statistical-principles guideline, was written to close that gap. Its central move is to separate the question from the statistical method used to answer it. First you define the target of estimation, the estimand. Only then do you choose an estimator and a sensitivity analysis. The published EMA guideline frames this as building "a common language" so that sponsors, regulators, and readers agree on what is being measured before arguing about how.

The five components of an estimand

The addendum specifies that an estimand has five attributes. Together they form a full sentence describing the effect.

1. Treatment

The treatment condition of interest and the alternative it is compared against. This sounds trivial, but it includes questions such as whether "treatment" means the drug alone or the drug plus permitted background therapy, and over what duration.

2. Population

The specific group of patients the clinical question is about, defined by the trial's eligibility criteria or by a clinically meaningful subgroup.

3. Variable (endpoint)

The measurement obtained from each patient that addresses the question, for example blood pressure at 24 weeks, time to a cardiovascular event, or a symptom score.

4. Intercurrent events and the strategy for them

Intercurrent events are things that happen after randomization and affect either the existence or the interpretation of the endpoint: stopping treatment, switching to rescue medication, or death. This is the component the framework added, and it is where most of the interpretive weight sits.

5. Population-level summary

How individual results are combined into a single comparison, such as a difference in means, an odds ratio, or a hazard ratio.

Change any one of these and you have changed the estimand, and therefore the meaning of the result.

Why intercurrent-event strategy is the pivot

The framework describes several distinct strategies for handling an intercurrent event, and each answers a subtly different clinical question. A recent perspective in Statistics in Medicine by Fleming and colleagues walks through them and argues that they are far from interchangeable.

Under a treatment policy strategy, the endpoint is collected and analyzed no matter what happens afterward, including after a patient stops the drug or starts something else. This preserves the randomized comparison and reflects the effect of a treatment decision as it plays out in the real world, which is why the authors treat it as the default that best protects against bias.

A hypothetical strategy instead estimates the effect that would have been seen in a world where the intercurrent event never occurred, for example if no one had needed rescue medication. That can be a legitimate question, but it relies on assumptions about an unobserved scenario that the data cannot fully verify.

A composite strategy folds the intercurrent event into the endpoint itself, so that, say, death or treatment discontinuation counts as a bad outcome. A while-on-treatment strategy only counts what happens up to the moment treatment stops. A principal stratum strategy restricts the question to the subgroup of patients defined by how they would respond to the event.

The Fleming perspective cautions that several of these strategies, particularly while-on-treatment, principal stratum, and some hypothetical constructions, can quietly break the protection that randomization provides, because they condition on things that happen after patients are randomized. When that happens, the groups being compared are no longer guaranteed to be balanced, and the estimate can be biased in a direction that is hard to predict. The authors favor designs that keep complete follow-up and analyze under a treatment policy strategy for exactly this reason.

How to read a result once you know the estimand

For anyone evaluating evidence, the practical lesson is to ask a single question before interpreting an effect size: which estimand produced this number? A drug can look more impressive under a hypothetical estimand, which imagines away the patients who could not tolerate it, than under a treatment policy estimand that counts everyone as randomized. Neither is dishonest; they answer different questions. The problem arises when a result generated under one estimand is discussed as though it answered another, which is how a real-world effectiveness claim can be built on an idealized number.

The estimand framework does not make any single trial "true." What it does is make the question explicit and auditable, so that the leap from a trial result to a clinical or policy claim can be inspected rather than assumed. That transparency is the point.

This article is educational and is not medical advice.