The ICH E9(R1) Estimand Framework Explained

Introduction

Clinical trials are often described as estimating a treatment effect, but the phrase "treatment effect" is much less precise than it sounds.

Consider a randomized trial comparing an experimental treatment with placebo. Suppose some participants discontinue treatment, some receive rescue medication, some die before the scheduled endpoint, and some remain on treatment through the final visit.

What exactly should the treatment effect represent?

Should it describe what would happen if everyone remained on their assigned treatment? Should patients who discontinue treatment still contribute their later outcomes? Should rescue medication be ignored? Should patients who discontinue because of lack of efficacy be considered treatment failures? Should the analysis estimate the effect of assignment to treatment or the effect of actually receiving treatment?

These are not merely statistical programming questions. They define the scientific question that the trial is intended to answer.

The ICH E9(R1) addendum introduced a structured framework for answering this problem through the concept of the estimand.

Why Was ICH E9(R1) Needed?

The original ICH E9 guideline established widely used principles for statistical analysis of clinical trials.

However, clinical trials became increasingly complex.

Patients may:

• Discontinue study treatment.
• Receive rescue medication.
• Switch treatment.
• Receive prohibited or additional therapy.
• Die before the planned assessment.
• Experience disease progression before the endpoint.
• Become otherwise unable to provide the planned measurement.

These events can occur after randomization and before the outcome of interest is observed.

Historically, statistical analysis plans often described these situations as "missing data" and then specified an imputation or modeling method.

ICH E9(R1) encourages a different way of thinking: first define what treatment effect is scientifically relevant, including how important post-randomization events should affect that treatment effect; then select an appropriate estimator and analysis method.

The Core Problem: What Treatment Effect Do We Want?

Suppose a trial compares treatment \(A\) with treatment \(B\), and the outcome is a change from baseline at Week 24.

A simplistic statement might be:

But this expression leaves several important questions unanswered.

What population is being considered? What exactly is treatment \(A\)? What happens if a participant discontinues treatment? What happens if rescue medication is administered? What outcome is being measured? What summary of the outcome is being compared?

The estimand framework forces these questions to be answered explicitly.

What Is an Estimand?

An estimand is the target of estimation.

It is not the statistical model. It is not the p-value. It is not the treatment difference observed in the dataset. It is not the estimator.

Instead, the estimand describes the precise treatment effect that the study intends to quantify.

The Five Attributes of an Estimand

ICH E9(R1) describes an estimand using five attributes:

These five attributes work together.

Changing any one of them can change the estimand.

Attribute 1: Treatment

The first attribute specifies the treatment conditions to be compared.

In a randomized parallel-group trial this might be:

versus:

But the treatment attribute can be more specific.

For example, it may specify:

• Assigned treatment
• A particular dose
• A dosing regimen
• A treatment strategy involving permitted concomitant therapy
• Treatment assignment regardless of adherence

The important point is that "treatment" in an estimand is a well-defined treatment condition rather than merely the name of a treatment arm.

Attribute 2: Population

The second attribute identifies the population for which the treatment effect is relevant.

For example:

• All randomized patients
• All randomized patients satisfying eligibility criteria
• Patients with a specific disease severity
• Patients with a particular biomarker status
• A prespecified subgroup

In a randomized trial, the target population should be distinguished from a particular analysis population used for computational convenience.

Attribute 3: Variable

The third attribute identifies the outcome variable.

Examples include:

• Change from baseline in systolic blood pressure at Week 12
• HbA1c at Week 24
• Time to cardiovascular death
• Overall survival
• Objective response by Week 24
• Number of exacerbations during follow-up
• Patient-reported quality-of-life score

The variable should be defined precisely enough that the target outcome is unambiguous.

For example, "clinical response" may be too vague. A stronger definition might specify the response criteria, assessment window, and relevant assessment rules.

Attribute 4: Intercurrent Events

This is one of the most important concepts in ICH E9(R1).

An intercurrent event is an event occurring after treatment initiation that affects either the interpretation or the existence of the measurement associated with the clinical question of interest.

Examples include:

• Permanent discontinuation of study treatment
• Initiation of rescue medication
• Treatment switching
• Use of prohibited medication
• Dose modification
• Death before the planned outcome assessment
• Other events that prevent or modify the planned outcome

The critical question is not simply: "How do we impute the missing value?"

The first question is: "What treatment effect should this event contribute to?"

Intercurrent Event vs. Missing Data

These concepts are related but are not interchangeable.

Suppose a participant stops treatment at Week 12 and therefore has no Week 24 measurement.

The treatment discontinuation is an intercurrent event. The absence of the Week 24 measurement is a missing-data issue.

The estimand framework asks how the treatment discontinuation should affect the scientific treatment effect. Only after that question has been addressed does the statistical analysis need to deal with any resulting missing observations.

Attribute 5: Population-Level Summary

The fifth attribute describes how the treatment effect is summarized at the population level.

Examples include:

• Difference in means
• Ratio of means
• Risk difference
• Risk ratio
• Odds ratio
• Hazard ratio
• Difference in survival probabilities at a fixed time
• Difference in median time to event

For a continuous endpoint, an estimand might use:

For a binary endpoint, it might use a risk difference:

The population-level summary completes the definition of the treatment effect.

The Five Attributes Together

A complete estimand can therefore be thought of as a structured statement:

The Five Intercurrent-Event Strategies

ICH E9(R1) describes several strategies for handling intercurrent events.

The principal strategies are:

Strategy 1: Treatment Policy

Under a treatment-policy strategy, the occurrence of the intercurrent event does not change the treatment effect being targeted.

The question is essentially: What is the effect of being assigned to the treatment strategy, regardless of whether the intercurrent event occurs?

For example, suppose patients may discontinue study treatment. A treatment-policy estimand may ask:

Patients who discontinue treatment remain relevant to the treatment-assignment effect.

This strategy is closely aligned with an intention-to-treat interpretation of treatment assignment, although an estimand should not simply be labeled "ITT" without specifying its attributes.

Strategy 2: Hypothetical

A hypothetical strategy asks what would happen under a specified hypothetical scenario.

For example:

This is a different scientific question from treatment policy.

The treatment-policy question includes treatment discontinuation in the treatment effect. The hypothetical question imagines a world in which the specified intercurrent event does not occur.

Strategy 3: Composite

Under a composite strategy, the intercurrent event becomes part of the endpoint itself.

Suppose the primary endpoint is clinical success at Week 24, but death before Week 24 is clinically important. A composite endpoint might define:

Death is therefore not treated merely as missing data. It is incorporated directly into the outcome definition.

Strategy 4: While on Treatment

The while-on-treatment strategy focuses on outcomes observed while patients remain on the assigned treatment.

For example:

This can be scientifically appropriate when the question is specifically about the pharmacologic or treatment effect during exposure.

However, it answers a narrower question than a treatment-policy estimand.

Strategy 5: Principal Stratum

The principal-stratum strategy defines the target population according to potential outcomes associated with the intercurrent event.

This is conceptually more difficult because the relevant subgroup may depend on potential outcomes that are not simultaneously observable.

For example, an investigator might want to compare treatment effects among patients who would remain free of rescue medication under either treatment condition.

That subgroup is defined using potential behavior under both treatment conditions.

Comparing the Intercurrent-Event Strategies

A Worked Example: HbA1c Trial

Consider a randomized Phase III diabetes trial comparing a new drug with placebo.

The primary endpoint is change from baseline in HbA1c at Week 24.

During the study:

• Some patients discontinue assigned treatment.
• Some patients receive rescue medication.
• Some patients remain on assigned treatment throughout.
• Some Week 24 measurements are unavailable.

A traditional SAP might simply state: "The primary analysis will use an MMRM under a missing-at-random assumption."

That describes an estimator and an assumption, but it does not fully describe the scientific question.

An estimand-first approach starts with the question.

Example A: Treatment-Policy Estimand

Suppose the scientific question is:

The five attributes can be described as:

The resulting estimand can be represented as:

where the potential outcomes are interpreted according to the treatment-policy strategy.

Example B: Hypothetical Estimand

Now suppose the scientific question is:

The treatment and population may remain unchanged. But the intercurrent-event strategy is different.

The two estimands therefore answer different scientific questions even though the endpoint and treatment comparison appear identical.

Example C: Composite Strategy

Suppose instead that receiving rescue medication is clinically considered treatment failure.

The endpoint could be redefined so that a patient who requires rescue therapy is classified as unsuccessful.

For example:

Now rescue medication is not something that must be removed through an imputation rule. It is part of the endpoint definition.

The Same Intercurrent Event Can Have Different Strategies

Consider treatment discontinuation.

The same event can be handled in several scientifically legitimate ways.

This is why saying simply "we handle discontinuations using MMRM" is not enough to define an estimand.

Estimand vs. Estimator vs. Estimate

These three terms are frequently confused.

For example, suppose the estimand is the difference in mean Week 24 change under a treatment-policy strategy.

An estimator might be an MMRM-based model-adjusted difference in means.

The resulting estimate might be:

The estimand is the target. The estimator is the method. The estimate is the number.

The Estimand Is Not the Same as the Analysis Method

This distinction is one of the most important lessons of ICH E9(R1).

Consider two analyses:

• MMRM
• Multiple imputation followed by ANCOVA

These are statistical methods.

They do not by themselves define the scientific treatment effect.

For example, two multiple-imputation analyses can target different estimands depending on how post-discontinuation outcomes are conceptualized and modeled.

Therefore:

Why "Missing at Random" Is Not an Estimand

Missing at random, or MAR, is a statistical assumption about the missing-data mechanism.

It does not specify:

• The treatment effect of interest
• The population
• The endpoint
• The intercurrent-event strategy
• The population-level summary

Therefore, MAR belongs to the estimation framework rather than serving as a replacement for an estimand.

Estimand and the Treatment-Policy Strategy

The treatment-policy strategy is especially important in randomized trials because it preserves the interpretation of treatment assignment.

Suppose:

• 100 patients are randomized to drug A.
• 100 patients are randomized to placebo.
• 20 patients in the drug A group discontinue treatment.
• 15 patients in the placebo group discontinue treatment.

A treatment-policy estimand does not redefine the treatment effect to apply only to patients who remain on treatment.

Instead, it asks about the effect of assignment to the treatment condition, with post-randomization events handled according to the specified strategy.

Estimand and Hypothetical Strategy

A hypothetical estimand is fundamentally counterfactual.

Suppose a patient receives rescue medication.

The hypothetical question may be:

What would the patient's outcome have been under treatment A in a world in which rescue medication did not occur?

This is not directly observed.

Consequently, hypothetical estimands often require stronger modeling assumptions than treatment-policy estimands.

Estimand and Composite Endpoints

Composite strategies can sometimes simplify the interpretation of clinically important events.

Suppose a cardiovascular trial measures time to cardiovascular death, but hospitalization is also clinically important.

A composite endpoint might define the outcome as time to:

• Cardiovascular death, or
• First cardiovascular hospitalization.

The hospitalization is then part of the endpoint rather than an event that must be handled separately after it occurs.

Estimand and While-on-Treatment Effects

A while-on-treatment estimand can be useful when the scientific question is specifically about the effect of treatment exposure.

For example:

This can be scientifically meaningful.

However, it should not be presented as though it were automatically the same as the treatment-policy effect.

Estimand and Principal Strata

Principal-stratum estimands are among the most conceptually challenging parts of the framework.

Suppose rescue medication is an intercurrent event. An investigator may want the treatment effect among patients who would not require rescue under either treatment.

Define the principal stratum:

where \(Y_A^{R}\) and \(Y_B^{R}\) indicate potential rescue behavior under the two treatment conditions.

The target treatment effect might then be:

The difficulty is obvious: for an individual patient, we generally observe only one treatment condition, so both potential rescue outcomes cannot be directly observed.

Consequently, identification and estimation may require strong assumptions.

Intercurrent Events Are Not All the Same

A common mistake is to create one blanket rule for all post-randomization events.

Instead, different intercurrent events may require different strategies.

The correct choice depends on the clinical question.

A Complete Estimand Specification

Consider a Phase III trial evaluating an investigational treatment for rheumatoid arthritis.

The primary endpoint is change from baseline in a disease activity score at Week 24.

Suppose the treatment-policy estimand is:

This is a substantially more informative specification than:

"The primary analysis will use MMRM."

From Estimand to Estimator

Once the estimand is defined, the statistical team selects an estimator.

For a continuous longitudinal endpoint, the estimator might use:

• MMRM
• ANCOVA
• Multiple imputation followed by an analysis model
• Other model-based approaches

The selected method must be appropriate for the estimand and the available data.

For example, an MMRM may estimate adjusted treatment differences at Week 24 using repeated observations over time.

PROC MIXED DATA=analysis;
  CLASS treatment visit subject;

  MODEL change =
    baseline
    treatment
    visit
    treatment*visit
    / DDFM=KR;

  REPEATED visit /
    SUBJECT=subject
    TYPE=UN;

  LSMEANS treatment*visit /
    DIFF CL;
RUN;

The model itself is not the estimand. It is an estimator used to obtain an estimate of the prespecified treatment effect.

Why the Same MMRM Can Support Different Questions

An MMRM can be used in analyses addressing different scientific questions, depending on how the data, intercurrent events, and assumptions are defined.

For example, one analysis may target a treatment-policy estimand. Another may be designed around a hypothetical estimand.

The difference lies in the scientific interpretation and assumptions, not simply in the name of the procedure.

Estimand and Missing Data

The estimand framework changes how missing data should be discussed.

Instead of starting with:

"We have 12% missing data. Which imputation method should we use?"

the analysis should start with:

"What treatment effect do we want to estimate when those post-randomization events occur?"

Then ask:

• What outcomes are observed?
• What outcomes are missing?
• What assumptions are needed?
• What estimator is appropriate?
• How sensitive are conclusions to those assumptions?

Estimand and Multiple Imputation

Multiple imputation is an estimation tool. It is not itself an estimand strategy.

For example, multiple imputation may be used to estimate a treatment-policy estimand if the imputation model is constructed to represent the intended post-intercurrent-event outcome distribution.

Likewise, reference-based multiple imputation can be used to construct certain hypothetical or other sensitivity scenarios.

The important question is not: "Are we using multiple imputation?"

The important question is: "What treatment effect does the imputation procedure attempt to estimate?"

Estimand and Sensitivity Analysis

ICH E9(R1) strongly reinforces the importance of examining the robustness of the primary conclusion.

A useful sensitivity-analysis program may consider:

• Alternative assumptions about missing outcomes
• Alternative handling of intercurrent events
• Alternative estimators
• Alternative covariance structures
• Reference-based assumptions
• Pattern-mixture models
• Selection models
• Tipping-point analyses

These analyses should be connected to clinically meaningful departures from the primary assumptions.

Estimand vs. Sensitivity Estimand

Not every sensitivity analysis needs to target exactly the same estimand.

An analysis can instead address a related but scientifically informative question.

For example:

This makes the sensitivity program much more interpretable than simply listing a collection of alternative statistical methods.

Estimand vs. Intercurrent Event Strategy

The intercurrent-event strategy is not a fifth statistical model. It is one of the five attributes defining the estimand.

This means that changing the strategy can change the target parameter.

For example:

where:

• \(\theta_{\mathrm{TP}}\) is a treatment-policy estimand.
• \(\theta_{\mathrm{Hyp}}\) is a hypothetical estimand.

They can use the same randomized treatment groups and the same nominal endpoint while answering different questions.

Worked Example: Rescue Medication

Suppose a Phase III trial evaluates pain reduction. The primary outcome is change in pain score at Week 12.

Patients may receive rescue analgesic medication.

There are at least three scientifically different questions.

Question 1: Treatment Policy

What is the treatment difference in Week 12 pain score regardless of whether patients use rescue medication?

This estimates the effect of treatment assignment in the presence of rescue therapy as it occurs in the trial.

Question 2: Hypothetical

What would the treatment difference have been if rescue medication had never been used?

This asks a counterfactual question.

Question 3: Composite

What is the treatment effect when patients requiring rescue medication are classified as treatment failures?

This changes the outcome definition.

All three questions can be scientifically legitimate. They are simply different estimands.

The Estimand Framework and Randomization

Randomization is fundamental to causal interpretation.

However, randomization does not automatically answer every post-randomization question.

Randomization provides a strong basis for comparing treatment groups under the randomized treatment conditions.

But once the question involves hypothetical post-randomization scenarios or principal strata, additional assumptions may be required.

Estimand and the Primary Objective

A well-written clinical trial objective should correspond closely to the primary estimand.

For example:

Weak objective: "To compare the efficacy of treatment A and placebo at Week 24."

This does not specify how treatment discontinuation or rescue medication affects the treatment effect.

Stronger objective: "To estimate the difference between treatment A and placebo in mean change from baseline in HbA1c at Week 24 among all randomized eligible patients, regardless of treatment discontinuation or rescue medication use."

The second statement is much closer to an estimand specification.

Estimands in the Protocol

The protocol should identify the treatment effect of interest before the statistical analysis is finalized.

A useful protocol section should describe:

• Primary estimand
• Key secondary estimands
• Treatment conditions
• Target population
• Primary outcome variable
• Important intercurrent events
• Strategy for each important intercurrent event
• Population-level summary

The statistical analysis plan can then describe how the estimand will be estimated.

Estimands in the Statistical Analysis Plan

A practical SAP structure is:

A Practical Estimand Table

A table like the following can make the estimand framework much easier to implement operationally.

One Trial Can Have Multiple Estimands

A modern clinical trial does not need to have only one estimand.

For example, a trial might define:

• One primary estimand
• Several key secondary estimands
• Additional estimands for clinically important treatment questions

For example:

This can provide a much richer description of treatment benefit than a single analysis.

Estimand vs. Per-Protocol Analysis

The estimand framework should not be reduced to a debate between "intention-to-treat" and "per-protocol."

Those terms describe analysis populations and strategies, whereas the estimand framework asks a more fundamental question: What causal treatment effect do we want?

A hypothetical or while-on-treatment estimand may resemble the scientific motivation behind a per-protocol analysis, but the concepts are not identical.

The estimand should be explicitly defined rather than inferred from an analysis-set label.

Estimand vs. As-Treated Analysis

Similarly, an as-treated analysis is not automatically an estimand.

An as-treated analysis changes how treatment exposure is represented in the analysis.

But that does not by itself specify:

• The target population
• The outcome
• The handling of intercurrent events
• The summary measure

The estimand framework therefore provides a more precise language.

Estimands and Time-to-Event Endpoints

The estimand framework applies to survival and time-to-event endpoints as well as continuous and binary endpoints.

Suppose the endpoint is overall survival.

The variable might be:

The population-level summary could be:

• Hazard ratio
• Difference in survival probability at Month 12
• Restricted mean survival time difference

These are different population-level summaries and therefore can correspond to different estimands.

Hazard Ratio Is Not the Only Possible Summary

Suppose:

Treatment:
    [Treatment A] versus [Treatment B]

Population:
    [Target population]

Variable:
    [Primary outcome and assessment time]

Intercurrent events:
    [Event 1] → [strategy]
    [Event 2] → [strategy]
    [Event 3] → [strategy]

Population-level summary:
    [Difference / ratio / hazard ratio / other summary]

PROC PHREG DATA=analysis;
  CLASS treatment;
  MODEL time*status(0) = treatment;
RUN;