Defining Intercurrent Events in Clinical Trials

Introduction

Clinical trials rarely proceed exactly as planned. Patients may discontinue treatment because of adverse events, start rescue medication, switch to another therapy, undergo a prohibited intervention, or die before the planned endpoint is observed.

These events occur after treatment assignment and can change what the treatment effect means.

For example, suppose a randomized trial compares Drug A with placebo using change from baseline in a continuous endpoint at Week 24.

A patient assigned to Drug A may discontinue treatment at Week 12 because of an adverse event but remain under follow-up and provide a Week 24 measurement.

Another patient may discontinue Drug A at Week 12 and immediately begin rescue therapy.

Both patients have experienced a treatment-related event after randomization, but the scientific question being answered may be different for each case.

Why Intercurrent Events Matter

A traditional analysis might simply say: "Analyze change from baseline to Week 24."

That sounds precise until patients experience events that alter the meaning of the Week 24 measurement.

Consider these scenarios:

These are not merely technical missing-data problems. They are questions about what treatment effect the trial is intended to estimate.

The Estimand Framework

The modern framework for handling these questions is the estimand framework described in ICH E9(R1).

An estimand specifies precisely what treatment effect is being targeted.

For a randomized clinical trial, an estimand generally describes five attributes:

1. The population of interest
2. The treatment condition being compared
3. The variable or endpoint of interest
4. How intercurrent events are handled
5. The population-level summary measure

The fourth component is particularly important because the same observed data can support different scientifically meaningful treatment effects depending on how intercurrent events are handled.

What Exactly Is an Intercurrent Event?

An intercurrent event is an event occurring after treatment initiation that affects either the interpretation of the endpoint or whether the endpoint can be observed as originally defined.

Examples include:

• Permanent treatment discontinuation
• Temporary treatment interruption
• Use of rescue medication
• Switching to another treatment
• Initiation of prohibited concomitant medication
• Receipt of an alternative intervention
• Death before the scheduled endpoint
• Organ transplantation or another event that changes the clinical trajectory

The exact list depends on the disease, endpoint, treatment, and scientific question.

Intercurrent Events Are Not the Same as Missing Data

This distinction is one of the most important concepts in estimand-based clinical trial analysis.

Suppose a patient discontinues study treatment at Week 12 but returns at Week 24 for the planned outcome assessment.

The treatment discontinuation is an intercurrent event. The Week 24 outcome is not necessarily missing.

Conversely, suppose the patient discontinues treatment and never returns for the Week 24 assessment.

Now there are two separate issues:

• The treatment discontinuation is an intercurrent event.
• The absent Week 24 measurement is missing data.

A Simple Example

Suppose a randomized trial compares Drug A with placebo. The endpoint is change in systolic blood pressure at Week 24.

Consider a patient assigned to Drug A:

What should happen to the Week 24 measurement?

There is no universally correct answer.

The answer depends on the estimand.

If the question is: "What is the effect of being assigned to Drug A regardless of whether patients remain on treatment?" then the Week 24 measurement may remain relevant.

If the question is: "What is the effect of Drug A while patients remain on Drug A?" then the post-discontinuation measurement may not answer the desired question.

The event itself therefore cannot be classified as simply "good" or "bad" for the analysis.

The Five ICH E9(R1) Strategies

ICH E9(R1) describes five principal strategies for handling intercurrent events:

Strategy 1: Treatment Policy

Under a treatment policy strategy, the outcome is considered regardless of whether the intercurrent event occurs.

For example, suppose some patients discontinue Drug A but remain under follow-up. The treatment-policy estimand may ask:

The treatment-policy strategy is closely aligned with an intention-to-treat interpretation when the outcome remains meaningful after the intercurrent event.

Example

Suppose the endpoint is change in HbA1c at Week 52. A patient assigned to Drug A switches to another glucose-lowering therapy at Week 30 but still has an HbA1c measurement at Week 52.

Under a treatment-policy strategy, the Week 52 value can still contribute to the treatment-policy estimand.

The question is not: "What would Drug A have done if the patient had remained on Drug A?"

Instead, it is closer to: "What is the effect of being assigned to Drug A, in the presence of the treatment changes that occur in practice?"

When Treatment Policy Is Attractive

Treatment policy can be particularly useful when the intercurrent event is part of the clinical reality that the treatment strategy is intended to influence.

For example:

• Patients may reasonably switch treatments in routine practice.
• Rescue therapy is clinically expected.
• The treatment effect of interest is an effect of treatment assignment.
• Follow-up after treatment discontinuation remains scientifically meaningful.

Strategy 2: Composite Strategy

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

Suppose the primary outcome is improvement in disease status. Instead of treating death as a separate problem, the endpoint could be defined as:

• Clinical improvement = success
• Death before assessment = failure

The endpoint is therefore a composite of the clinical outcome and the intercurrent event.

Example: Time-to-Event Endpoint

Suppose a cardiovascular trial evaluates time to clinical deterioration. Death may be incorporated directly into the endpoint:

In this situation, death is not treated merely as missing information. It is explicitly part of the endpoint definition.

Composite Strategy for Binary Endpoints

Suppose the original endpoint is response at Week 24. The composite endpoint could define:

This can produce a clinically interpretable endpoint when the intercurrent event represents an unequivocal unfavorable outcome.

Strategy 3: Hypothetical Strategy

A hypothetical strategy defines the treatment effect under a specified hypothetical scenario in which the intercurrent event does not occur.

For example:

or:

The hypothetical strategy is often scientifically appealing because it can address questions about treatment efficacy in an idealized situation.

However, it generally requires stronger assumptions than a treatment-policy strategy.

Why Hypothetical Strategies Are Challenging

Suppose a patient receives rescue medication at Week 16. There is no observed Week 24 outcome under the hypothetical scenario: "What would the Week 24 outcome have been if this patient had not received rescue medication?"

The hypothetical outcome must therefore be inferred using a statistical model or other assumptions.

This is fundamentally different from simply observing the patient's actual Week 24 value.

Strategy 4: While-on-Treatment Strategy

The while-on-treatment strategy focuses on outcomes occurring before the intercurrent event.

Suppose a patient discontinues treatment at Week 18. The estimand may ask: "What is the treatment effect while patients remain on assigned treatment?"

Measurements after Week 18 are not used to define the target treatment effect.

For a longitudinal endpoint:

where \(T_{\text{ICE}}\) is the time of the relevant intercurrent event.

Example

A drug is administered continuously for 24 weeks, and investigators want to understand its effect while patients actually receive the drug.

A patient discontinues at Week 16. The while-on-treatment strategy uses the information accumulated through the time of discontinuation rather than interpreting later observations as part of the on-treatment effect.

While-on-Treatment Is Not Automatically Per-Protocol

These concepts are related but should not be treated as identical.

A while-on-treatment estimand is a precisely defined scientific target. A per-protocol analysis is a particular analytic approach involving adherence or protocol-compliance criteria.

A protocol can contain a while-on-treatment estimand without simply equating it to "analyze only protocol-compliant patients."

Strategy 5: Principal Stratum

The principal stratum strategy defines the treatment effect within a subgroup characterized by whether an intercurrent event would occur under particular treatment conditions.

This is conceptually different from simply conditioning on an event that has already been observed after treatment assignment.

For example, investigators might be interested in the treatment effect among patients who would not require rescue medication under either treatment.

The relevant subgroup is a latent principal stratum because, for an individual patient, we generally cannot observe what would happen under both treatment conditions.

A Side-by-Side Comparison of the Five Strategies

One Intercurrent Event Can Have Different Strategies

There is no requirement that all intercurrent events in a study use the same strategy.

For example:

The correct choice depends on the clinical question.

Example: Rescue Medication

Suppose a trial evaluates a treatment for chronic pain. The primary endpoint is change in pain score at Week 12.

Patients are allowed to receive rescue analgesia if pain becomes severe.

The investigators must decide what the Week 12 estimand means.

Option A: Treatment Policy

Use the Week 12 pain score regardless of rescue medication.

This answers a pragmatic question about treatment assignment in the presence of rescue therapy.

Option B: Composite

Define rescue medication use as treatment failure.

A patient who requires rescue medication is therefore classified as a failure, possibly regardless of the subsequent pain measurement.

Option C: Hypothetical

Estimate what the Week 12 pain score would have been if the patient had not received rescue medication.

Option D: While on Treatment

Use information before rescue medication occurs and define the target around the patient's treatment experience before the event.

These are four different estimands.

Example: Treatment Discontinuation

Suppose a patient permanently discontinues treatment because of an adverse event at Week 8.

There are several possible scientific questions.

Example: Treatment Switching

Treatment switching can be particularly important in oncology trials.

Suppose a patient randomized to Drug A progresses and subsequently receives Drug B.

The investigator must decide whether the post-switch outcome belongs to the target treatment effect.

A treatment-policy estimand may include outcomes after switching.

A while-on-treatment estimand may stop the treatment-specific outcome assessment at switching.

A hypothetical estimand might ask what would have happened had the patient not switched.

A principal-stratum estimand might target patients who would not switch under specified treatment conditions.

Example: Death

Death requires particularly careful consideration because the event may make the originally defined outcome impossible to observe.

Suppose the endpoint is quality of life at Week 24. A patient who dies at Week 16 cannot provide a Week 24 quality-of-life score.

The analysis cannot simply assume that the Week 24 value is an ordinary missing observation without considering the estimand.

Composite Approach

Death can be incorporated into the endpoint. For example, a composite endpoint might rank death as worse than any observed quality-of-life value.

Treatment Policy

If the treatment-policy question concerns the patient's status at Week 24, death may require a specific definition of the outcome rather than simply ignoring the patient.

Hypothetical Approach

A hypothetical estimand could ask about quality of life at Week 24 under a scenario in which the patient had not died.

Such a question requires particularly strong assumptions because the hypothetical outcome is not observable.

Intercurrent Events and Potential Outcomes

The estimand framework can be understood using potential outcomes. Suppose:

denotes the outcome patient \(i\) would experience under treatment 1, and:

denotes the outcome under treatment 0.

An intercurrent event complicates this notation because the event may itself depend on treatment.

For example, rescue medication may have potential outcomes:

and the clinical outcome may depend on whether rescue medication occurs.

Thus, the treatment effect may need to be defined jointly with the event strategy.

Why Randomization Does Not Automatically Solve the Problem

Randomization protects the comparison between treatment groups at baseline.

It does not automatically answer what should happen after a post-randomization event.

For example, conditioning only on patients who remain on treatment can create selection problems because treatment discontinuation may be related to prognosis, treatment response, or adverse events.

Therefore:

The estimand must still be explicitly defined.

Estimand vs. Estimator vs. Estimate

These three terms are often confused.

For example, an estimand might be:

A statistical estimator might be an adjusted mean difference from a prespecified ANCOVA model.

The estimate might then be:

Changing the handling of an intercurrent event can change the estimand even if the same statistical model is used.

Why the Estimand Comes Before the Analysis

Consider two analysts who both fit an MMRM.

Analyst A intends to estimate a treatment-policy estimand. Analyst B intends to estimate a hypothetical estimand assuming no rescue medication.

Even though both use an MMRM, they are not necessarily estimating the same quantity.

The differences may appear in:

• Which observations are included
• How rescue medication is handled
• How post-event outcomes are interpreted
• The missing-data assumptions
• The imputation model
• The sensitivity analyses

Intercurrent Events and Missing Data

Once the estimand has been defined, missing data may still remain.

For example:

The intercurrent event and missingness have therefore been separated.

Common Intercurrent Events

Not Every Protocol Deviation Is an Intercurrent Event

A protocol deviation and an intercurrent event are not synonymous.

A protocol deviation describes a departure from the planned conduct of the study. An intercurrent event describes a post-randomization event that affects the interpretation of the outcome relevant to the estimand.

Some events can be both. Others may be one but not the other.

A Practical Event-Classification Table

A useful protocol-development exercise is to create an intercurrent-event matrix.

How to Define an Intercurrent Event Precisely

A vague statement such as: "Patients who discontinue treatment will be handled appropriately." is not sufficient for a rigorous estimand.

Instead, specify:

1. The event definition.
2. The time window in which it qualifies.
3. Whether temporary and permanent events are distinguished.
4. Whether the event changes the endpoint definition.
5. The estimand strategy applied to the event.
6. How post-event observations are treated.
7. How missing observations are subsequently handled.
8. The statistical method used to estimate the resulting estimand.

Example of a Precise Definition

Instead of:

"Rescue medication will be considered in the analysis."

a more precise definition might state:

The exact wording should be adapted to the clinical setting and protocol.

Estimand Matrix for a Clinical Trial

A useful way to organize the primary analysis is to create a matrix linking each important intercurrent event to its estimand strategy.

This type of table makes the estimand transparent to statisticians, clinicians, regulators, and other members of the study team.

Estimand and the Primary Hypothesis

Suppose the treatment effect is defined as a difference in mean change from baseline:

The meaning of \(Y(1)\) and \(Y(0)\) depends on how intercurrent events are handled.

Under treatment policy, \(Y(1)\) may represent the endpoint regardless of post-randomization treatment changes.

Under a hypothetical strategy, \(Y(1)\) may represent the outcome under a specified hypothetical scenario in which an event does not occur.

Under a while-on-treatment strategy, \(Y(1)\) may represent the endpoint defined only before the event.

Thus the mathematical expression alone does not completely define the estimand.

Binary Endpoint Example

Suppose the primary endpoint is response at Week 24. Let:

and:

A risk difference might be:

But the definition of response must also specify what happens when an intercurrent event occurs.

For example, under a composite strategy:

The event strategy is therefore part of the endpoint definition.

Time-to-Event Endpoints

Intercurrent events can be particularly important in survival analyses.

Suppose the endpoint is progression-free survival. A treatment switch after progression may not be treated in the same way as a treatment switch before progression.

Similarly, death may be part of the endpoint rather than a source of missing data.

For a time-to-event estimand, the analyst should define:

• The event of interest
• The time origin
• The censoring rules
• The role of treatment switching
• The role of rescue treatment
• The role of competing events
• The treatment-policy or hypothetical interpretation

Intercurrent Events and Censoring

One common mistake is to automatically censor a patient at the time of an intercurrent event.

That is not necessarily appropriate.

Censoring is an analytical operation. The intercurrent event is a clinical event.

Whether the event should lead to censoring depends on the estimand and statistical analysis.

Intercurrent Events in Repeated-Measures Trials

Suppose patients have measurements at:

• Baseline
• Week 4
• Week 8
• Week 12
• Week 24

A patient discontinues treatment at Week 12.

The Week 4, Week 8, and Week 12 measurements may still be observed. The Week 24 measurement may also be observed, depending on follow-up.

The estimand determines which of these observations are relevant to the target treatment effect.

MMRM Does Not Define the Estimand

Mixed models for repeated measures are frequently used in clinical trials, but an MMRM is a statistical model rather than an estimand.

For example, an MMRM may use all available post-baseline observations under a missing-at-random assumption.

That does not by itself establish whether the resulting treatment effect represents:

• A treatment-policy estimand
• A hypothetical estimand
• A while-on-treatment estimand

The connection must be established explicitly.

Multiple Imputation Does Not Define the Estimand

The same principle applies to multiple imputation.

Multiple imputation is a method for handling missing values.

It does not answer the scientific question of what should happen after an intercurrent event.

For example, a hypothetical estimand may require imputation under assumptions about what would have happened without rescue medication. That is an estimand-driven use of imputation.

By contrast, simply imputing all missing Week 24 values with a standard multiple-imputation model does not automatically make the analysis hypothetical.

Sensitivity Analyses

Because intercurrent-event handling often depends on assumptions, sensitivity analyses are important.

For example, the primary estimand might use a treatment-policy strategy while a sensitivity analysis evaluates a hypothetical strategy.

Alternatively, the primary analysis might use a composite strategy while alternative assumptions are examined.

The goal is not to create an arbitrary collection of analyses. The goal is to evaluate the robustness of the clinical conclusion to reasonable alternative assumptions.

Estimand Strategy vs. Sensitivity Analysis

A useful distinction is:

A Worked Example: Week 24 Continuous Endpoint

Consider a randomized Phase III trial comparing Drug A with placebo. The primary endpoint is change from baseline in a continuous disease score at Week 24.

Suppose the study team identifies three important intercurrent events:

Step 1: Define the Population

Suppose the target population is all randomized patients eligible for the primary analysis.

This defines who the treatment effect applies to.

Step 2: Define the Treatment Conditions

The comparison is:

Step 3: Define the Variable

The endpoint is:

Step 4: Define the Intercurrent-Event Strategies

Suppose the team decides:

Now the estimand is substantially more precise.

Step 5: Define the Summary Measure

Suppose the treatment effect is the difference in mean Week 24 change:

The five components together define the treatment effect much more clearly than simply saying "compare mean change at Week 24."

A Full Estimand Statement

A practical estimand statement might read conceptually as:

The exact language should be adapted to the protocol and endpoint.

Why "Treatment Discontinuation = Missing" Is Wrong

A common historical approach is:

This sequence skips the estimand question.

A patient can discontinue treatment and still have a clinically relevant post-discontinuation measurement.

Alternatively, a patient can remain on treatment but have a missing endpoint assessment.

Therefore:

Common Mistakes in Defining Intercurrent Events

1. Calling every discontinuation missing data. Discontinuation is an event; missingness is a separate issue.
2. Choosing the statistical model before defining the estimand. The analysis should be selected to estimate the scientific target.
3. Assuming treatment policy means "ignore everything." The event remains clinically relevant; the estimand simply incorporates outcomes regardless of the event.
4. Automatically censoring at treatment discontinuation. Censoring is not the default consequence of every intercurrent event.
5. Calling a per-protocol analysis a principal-stratum analysis. The concepts are not interchangeable.
6. Using hypothetical analyses without explaining assumptions. The unobserved hypothetical outcome requires identification assumptions.
7. Ignoring death. Death can fundamentally change whether the original endpoint exists.
8. Using one strategy for every event. Different intercurrent events can require different strategies.
9. Failing to define the event itself. "Rescue therapy" must be operationally defined.
10. Changing the strategy after observing results. The primary estimand should be prespecified.

Intercurrent Events and the Statistical Analysis Plan

The SAP should translate the estimand into reproducible analysis rules.

For every important intercurrent event, specify:

• Operational event definition
• Date/time definition
• Relevant treatment exposure
• Whether the event is included in the primary estimand
• Selected intercurrent-event strategy
• Handling of post-event observations
• Handling of missing observations
• Primary statistical model
• Sensitivity analyses

Intercurrent Event Documentation Template

A Practical Workflow

How to Think About the Five Strategies

A useful memory aid is to translate each strategy into a question.

Choosing Between Strategies

The choice should be driven by the clinical question rather than by statistical convenience.

Choose Treatment Policy When:

• The effect of treatment assignment is the main question.
• Post-event outcomes remain clinically meaningful.
• Treatment switching or rescue reflects real-world treatment behavior.

Choose Composite When:

• The intercurrent event represents an important clinical outcome.
• Event occurrence should count as treatment failure or another endpoint state.
• The combined endpoint has a coherent clinical interpretation.

Choose Hypothetical When:

• The scientific question explicitly concerns an unrealized treatment scenario.
• The no-event scenario is clinically meaningful.
• The assumptions required for estimation are acceptable and can be evaluated.

Choose While on Treatment When:

• The scientific interest is specifically in treatment exposure before the event.
• Post-event outcomes no longer represent the desired treatment effect.

Choose Principal Stratum When:

• The treatment effect within a well-defined event-related subgroup is scientifically important.
• The necessary assumptions and identification strategy are defensible.

Primary and Secondary Estimands

A clinical trial can have more than one estimand.

For example, a study might define:

This can provide complementary information without pretending that all analyses answer exactly the same question.

Estimand Alignment Across the Trial

A strong clinical trial has alignment between:

Misalignment at any point can make the final interpretation ambiguous.

Example of Misalignment

Suppose the protocol says the primary question is: "What is the effect of treatment assignment regardless of rescue medication?"

That suggests a treatment-policy estimand.

However, the SAP censors patients at rescue medication and analyzes only pre-rescue observations.

That analysis may no longer estimate the stated treatment-policy estimand.

This is why estimand and analysis should be reviewed together.

Intercurrent Events and Regulatory Communication

Explicit estimand definitions improve communication among:

• Clinical investigators
• Biostatisticians
• Clinical pharmacologists
• Data managers
• Medical writers
• Regulatory reviewers
• Other members of the development team

Instead of debating whether a patient should be "excluded," the team can ask a more precise question: "Which treatment effect are we trying to estimate, and what strategy does that imply for this event?"

Intercurrent Events in SAP Programming

Statistical programmers should translate the estimand into explicit analysis variables and derivation rules.

For example, if rescue medication is a composite failure:

if rescue_medication = 1
   then response = 0;

If death is also a composite failure:

if death_before_week24 = 1
   then response = 0;

The exact implementation depends on the endpoint and SAP.

For continuous endpoints, programmers may instead create event indicators, analysis windows, censoring variables, or imputation flags.

ADaM and Intercurrent Events

In a CDISC environment, intercurrent-event handling may require variables that allow the analysis dataset to distinguish:

• Treatment discontinuation dates
• Rescue medication dates
• Treatment switching dates
• Death dates
• Endpoint assessment dates
• Post-event versus pre-event observations
• Analysis flags
• Composite endpoint indicators

The exact implementation should be driven by the analysis specification.

An analysis-ready dataset should make it possible to reconstruct the prespecified estimand without relying on undocumented programming decisions.

Common Programming Pitfall

Suppose a patient has:

A programmer should not simply decide that the Week 24 value is "invalid." The correct handling depends on the estimand.

For a treatment-policy estimand, the Week 24 measurement may remain relevant. For a composite strategy, rescue medication may define failure regardless of the Week 24 score. For a while-on-treatment estimand, the Week 12 value may be the final relevant observation. For a hypothetical estimand, the observed post-rescue values may not represent the desired hypothetical outcome.

Intercurrent Events and Data Review

Clinical data review should identify intercurrent events before database lock.

Useful review variables include:

• Date of treatment discontinuation
• Reason for discontinuation
• Date of rescue therapy
• Type of rescue therapy
• Date of treatment switch
• Date of death
• Post-event endpoint measurements
• Whether the endpoint remains meaningful after the event

This allows the statistical team to verify that the analysis population and endpoint derivations match the prespecified estimand.

Intercurrent Events and Blinded Review

Whenever possible, event classifications and derivation rules should be reviewed before treatment groups are unblinded.

Examples include:

• Confirming rescue medication definitions
• Confirming treatment-switch dates
• Confirming composite endpoint rules
• Confirming analysis windows
• Reviewing ambiguous event classifications

This reduces the risk that post-unblinding decisions inadvertently depend on observed treatment effects.

Why Event Definitions Need Operational Rules

Consider the term "rescue medication."

Different investigators might interpret this as:

• Any additional medication
• Only protocol-listed medications
• Only medication above a specified dose
• Medication used after inadequate response
• Medication used for safety reasons

Therefore, the protocol must define exactly what qualifies.

Intercurrent Event Timing Matters

The same event may have different implications depending on when it occurs.

For example, rescue medication at:

• Week 2
• Week 12
• Week 23

may have different clinical implications for a Week 24 endpoint.

A rigorous definition therefore includes a time window when appropriate.

Temporary vs. Permanent Discontinuation

Treatment interruption and permanent discontinuation should not automatically be treated as the same event.

For example:

The estimand should specify whether these events are combined or treated separately.

What the Statistical Analysis Should Report

The final clinical study report should make the handling of important intercurrent events understandable.

Depending on the study, reporting may include:

• Frequency of treatment discontinuation
• Reasons for discontinuation
• Frequency of rescue medication use
• Treatment switching
• Deaths
• Post-event measurements
• Primary estimand definition
• Primary analysis
• Sensitivity analyses

This allows readers to understand not only the numerical treatment effect but also the clinical question that effect answers.

A Compact Estimand Checklist

A Practical Estimand Workflow for Biostatisticians

The Most Important Concept

The most important idea is simple: An intercurrent event is not merely a nuisance that must be removed from the dataset.

It is a clinical event that forces the investigator to clarify what treatment effect the trial is actually intended to estimate.

Consider treatment discontinuation.

The question is not automatically: "What do we do with this patient?"

The better question is: "What treatment effect do we want the patient's experience to contribute to?"

Once that question is answered, the appropriate intercurrent-event strategy and statistical analysis become much easier to define.

Key Takeaways

• An intercurrent event is not synonymous with missing data.
• Treatment discontinuation does not automatically imply censoring or exclusion.
• The estimand defines the scientific question; the analysis estimates that question.
• The five principal strategies are treatment policy, composite, hypothetical, while on treatment, and principal stratum.
• Different intercurrent events can use different strategies within the same trial.
• Death requires explicit consideration because the original endpoint may cease to exist after death.
• Hypothetical estimands require assumptions about outcomes that cannot actually be observed under the specified scenario.
• Principal-stratum estimands can involve latent subgroups and substantial identification assumptions.
• MMRM, multiple imputation, censoring, and other statistical methods do not define the estimand by themselves.
• The primary estimand and its intercurrent-event strategies should be prespecified before unblinding.

References

International Council for Harmonisation (ICH) (2019). E9(R1): Addendum on Estimands and Sensitivity Analysis in Clinical Trials to the Guideline on Statistical Principles for Clinical Trials.

International Council for Harmonisation (ICH) (1998). E9: Statistical Principles for Clinical Trials.

ICH E9(R1) describes the estimand framework and the principal strategies for addressing intercurrent events in clinical trials, including treatment policy, composite, hypothetical, while-on-treatment, and principal-stratum strategies.

European Medicines Agency (EMA). ICH E9(R1) Addendum on Estimands and Sensitivity Analysis in Clinical Trials.

U.S. Food and Drug Administration (FDA). E9(R1) Statistical Principles for Clinical Trials: Addendum — Estimands and Sensitivity Analysis.