ICH E9(R1) Estimands: Practical Implementation

Introduction

One of the most important changes introduced by ICH E9(R1) is a shift in how clinical-trial treatment effects are specified. Instead of starting with an endpoint and an analysis model, the framework asks a more fundamental question: What treatment effect do we actually want to estimate?

That question becomes especially important when events occur after treatment initiation that affect the interpretation or existence of outcome measurements. Examples include treatment discontinuation, rescue medication, treatment switching, death, transplantation, or other events that change what the observed outcome means.

ICH E9(R1) calls these intercurrent events. They are not simply inconvenient missing-data mechanisms. The way an intercurrent event is handled is part of the clinical question itself.

Why ICH E9(R1) Matters

Traditional trial descriptions often say something such as: “Compare treatment groups on change from baseline at Week 24 using an ANCOVA model.”

That statement identifies an outcome, a time point, and an analysis method, but it may leave a critical question unanswered: What should happen to the treatment effect definition if a patient stops randomized treatment and starts rescue therapy before Week 24?

Consider two patients with the same Week 24 value:

• Patient A remained on randomized treatment throughout follow-up.
• Patient B discontinued randomized treatment at Week 8 and received rescue medication.

The same observed Week 24 measurement can answer different clinical questions depending on whether the treatment-policy question, a hypothetical question, or a while-on-treatment question is being asked.

ICH E9(R1) provides a structured framework so that this distinction is made before the data are analyzed.

Estimand, Estimator, and Estimate

These three terms are related but should not be used interchangeably.

A useful conceptual sequence is:

The Five Attributes of an ICH E9(R1) Estimand

ICH E9(R1) describes five attributes that together define the treatment effect of interest:

1. Population: the patients targeted by the treatment-effect question.
2. Treatment condition: the treatment conditions being compared.
3. Variable: the outcome used to quantify the patient response.
4. Intercurrent events: how events after treatment initiation are reflected in the treatment-effect question.
5. Population-level summary: how individual outcomes are aggregated into the treatment effect.

Attribute 1: Population

The population attribute identifies the population for which the treatment effect is relevant. This is more specific than simply naming the randomized population.

A protocol might define the population as:

The estimand should make clear whether the target is the full randomized population, a subset defined by eligibility, or another clinically justified population.

In a randomized trial, the population attribute should generally be aligned with the trial's causal question and analysis population. It should not be silently changed after looking at the observed data.

Attribute 2: Treatment Condition

The treatment attribute specifies the treatment conditions being compared. This can include:

• Experimental treatment versus placebo
• Experimental treatment versus active comparator
• Different doses or regimens
• Treatment strategies involving treatment switching

A precise description matters when dose, schedule, background therapy, or treatment duration is part of the clinical question.

Attribute 3: Variable

The variable is the outcome used to quantify patient response.

Examples include:

• Change from baseline in a continuous endpoint at Week 24
• Responder status at Week 24
• Time to first disease progression or death
• Overall survival
• A repeated patient-reported outcome score
• A composite clinical outcome

The variable should be operationally defined. “Clinical response” is usually too vague. A stronger specification identifies the measurement, timing, derivation, and rules for events that affect its interpretation.

Attribute 4: Intercurrent Events

An intercurrent event occurs after treatment initiation and affects either the interpretation or the existence of the outcome measurement relevant to the clinical question.

Examples include:

Attribute 5: Population-Level Summary

The final attribute specifies how patient-level outcomes are summarized.

Examples include:

• Difference in means
• Ratio of means
• Risk difference
• Risk ratio
• Odds ratio
• Difference in restricted mean survival time
• Hazard ratio, when appropriate to the estimand and model
• Another prespecified population-level contrast

The summary measure is not a cosmetic reporting choice. Changing it can change the estimand.

Intercurrent Event Strategies

ICH E9(R1) describes several strategies for specifying how an intercurrent event is handled.

The five commonly discussed strategies are:

Treatment-Policy Strategy

Under a treatment-policy strategy, the clinical question asks what happens regardless of whether the intercurrent event occurs.

For example:

This is often aligned with an effect of treatment assignment question in which events that occur after randomization are part of what happens under each treatment strategy.

For a continuous endpoint, the estimand might be expressed as:

where \(Y_{24}(A)\) and \(Y_{24}(B)\) represent the potential Week 24 outcomes under assignment to the respective treatment conditions, with the treatment-policy handling of intercurrent events built into the question.

Practical implication

The data-collection plan should try to obtain outcome information after treatment discontinuation or rescue when that information is part of the treatment-policy estimand. Simply discarding those measurements can move the analysis away from the stated question.

Hypothetical Strategy

A hypothetical strategy asks what the treatment effect would have been under a specified scenario in which the intercurrent event did not occur.

For example:

This is a different clinical question from the treatment-policy estimand.

For example:

The exact counterfactual scenario must be defined carefully. “Hypothetical” is not a license to remove inconvenient observations without a scientific rationale.

Composite Strategy

Under a composite strategy, the intercurrent event is incorporated directly into the outcome.

For example, suppose a clinical trial studies a binary response at Week 24 and considers death before Week 24 clinically important. A composite endpoint could define:

• Responder if the patient meets the response criterion and remains alive through Week 24.
• Non-responder if the patient dies before Week 24.

The estimand then concerns the probability of the composite outcome, not simply the probability of the original response among patients with an observed Week 24 measurement.

While-on-Treatment Strategy

A while-on-treatment strategy defines the treatment effect using outcomes while patients are exposed to the assigned treatment, according to a prespecified treatment-exposure window.

For example:

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

The exposure window must be operationalized. A vague statement such as “while on treatment” is insufficient for a production SAP.

Principal-Stratum Strategy

A principal-stratum strategy defines the treatment effect within a subgroup characterized by post-randomization potential event behavior.

For example, a scientific question might concern treatment effects among patients who would not require rescue under either treatment condition.

Conceptually, a principal stratum can be written as a subgroup defined by potential intercurrent-event outcomes:

where \(R_i(A)\) and \(R_i(B)\) denote whether patient \(i\) would experience the specified event under each treatment condition.

The Same Intercurrent Event Can Support Different Estimands

Consider rescue medication at Week 12.

The same event can lead to different questions:

The event has not changed. The clinical question has changed.

Estimand Is Not the Same as “Handling Missing Data”

This is one of the most important implementation concepts.

Suppose a patient discontinues treatment at Week 12 and has no Week 24 measurement.

There are two distinct questions:

1. How should treatment discontinuation be reflected in the estimand?
2. Given the selected estimand, how should the unobserved Week 24 outcome be handled statistically?

The first is an estimand question. The second is an estimation question.

A Worked Example: Continuous Endpoint

Suppose a randomized trial compares Treatment A with placebo on change from baseline in a continuous symptom score at Week 24.

The estimand is therefore:

with the treatment-policy interpretation of intercurrent events.

Worked Example: Hypothetical Estimand

Now suppose the scientific question changes:

The variable and treatment groups have not changed. The intercurrent-event strategy has changed, so the estimand has changed.

The analysis may now require a model or other method capable of estimating a counterfactual outcome under the no-rescue/no-discontinuation scenario.

Worked Example: Composite Estimand

Suppose the endpoint is a responder definition, but death before Week 24 is considered a clinically important failure.

Instead of treating death as a missing Week 24 response, define the outcome as a composite:

The treatment effect is then a contrast in the probability of this composite outcome.

Estimand Versus Endpoint

An endpoint is a measurement or outcome definition. An estimand is a complete treatment-effect specification.

For example, “change from baseline in blood pressure at Week 12” describes an endpoint. It does not fully answer:

• Who is the target population?
• What happens if treatment is discontinued?
• What happens if rescue medication is initiated?
• What treatment contrast is being made?
• How are individual outcomes summarized?

The estimand adds those elements.

Estimand Versus Analysis Set

A common implementation error is to define the estimand after defining the analysis set.

For example:

Restricting to patients with observed Week 24 values can condition on post-randomization behavior and may answer a different question from the randomized treatment-effect question.

Estimand-Driven Trial Design

ICH E9(R1) emphasizes that estimands can influence trial design and conduct.

Suppose the primary estimand uses a treatment-policy strategy and the endpoint is measured at Week 52.

Then collecting outcome information after treatment discontinuation may be essential because those outcomes are part of the treatment-policy question.

By contrast, a hypothetical estimand may require a different data structure, assumptions, or modeling strategy to estimate the counterfactual outcome.

Estimands and Randomization

Randomization identifies a causal comparison between treatment assignments under suitable assumptions, but the estimand still determines exactly which causal question is being asked.

For a simple treatment-policy estimand, the treatment contrast can often be interpreted as the effect of assignment to the treatment strategies, including what happens after treatment discontinuation or rescue as specified by the estimand.

For hypothetical or principal-stratum estimands, additional assumptions may be needed because the target concerns counterfactual outcomes or post-randomization strata.

Choosing the Estimator

Once the estimand is defined, the estimator should be selected to target it.

For a continuous endpoint, candidate approaches may include:

• Analysis of covariance
• Repeated-measures models
• Mixed models for repeated measures
• Multiple imputation followed by an analysis model
• Other likelihood-based or semiparametric approaches

For binary outcomes, possible estimators include:

• Risk difference
• Risk ratio
• Odds ratio
• Regression-adjusted treatment contrasts

For time-to-event outcomes, possible estimands and estimators can involve:

• Survival probabilities at a fixed time
• Restricted mean survival time
• Hazard-based contrasts
• Other prespecified survival-function summaries

The choice should be driven by the estimand and the scientific question, not by the availability of a familiar software procedure.

A Useful Mathematical Distinction

Let \(\theta\) denote the estimand.

An estimator can be represented abstractly as:

where \(Y\) denotes observed outcomes, \(A\) treatment assignment, and \(X\) any prespecified covariates or other analysis information.

The exact estimator may be much more complicated, but the conceptual point is simple: \(\theta\) is the target; \(\widehat{\theta}\) is the estimate produced by the chosen procedure.

Sensitivity Analysis

Sensitivity analyses investigate whether the primary conclusion is robust to assumptions that are not fully verified by the observed data.

For an estimand-focused analysis, sensitivity analyses should be connected to the assumptions required to estimate the chosen estimand.

Examples include:

• Alternative assumptions about missing outcomes
• Alternative models for the outcome process
• Alternative assumptions about treatment discontinuation
• Alternative handling of rescue medication
• Alternative covariance structures
• Pattern-mixture approaches
• Tipping-point analyses
• Reference-based imputation approaches where scientifically appropriate

Primary Analysis and Sensitivity Analysis Should Not Contradict the Estimand

Suppose the primary estimand is treatment policy, but the primary analysis simply excludes patients after treatment discontinuation.

That can create a mismatch between the estimand and estimator.

The question should instead be:

If the answer is no, the analysis should be reconsidered or the estimand should be reconsidered.

Missing Data Under a Treatment-Policy Estimand

Suppose a patient discontinues treatment at Week 16 but remains willing to provide outcome measurements through Week 24.

Under a treatment-policy estimand, those measurements can remain relevant because the question concerns outcomes regardless of discontinuation.

If the patient also stops providing measurements, missing-data methods are then required to address the unobserved outcome.

Thus:

They are related but distinct components of the analysis.

Example: Rescue Medication

Suppose patients can receive rescue medication after Week 8.

Three possible estimands might be:

None of these is inherently “the correct way” without reference to the clinical objective.

Example: Treatment Discontinuation Due to Adverse Event

Treatment discontinuation can have different meanings depending on why it occurred.

A patient may discontinue because:

• The treatment is ineffective.
• The treatment causes an adverse event.
• The patient chooses to stop treatment.
• The investigator recommends discontinuation.
• The protocol requires discontinuation.

An estimand can treat these events collectively or separately if the clinical question requires it.

For example, a composite strategy might define discontinuation for lack of efficacy as treatment failure, whereas a treatment-policy strategy might continue to consider later outcomes regardless of discontinuation.

Example: Death

Death is a particularly important intercurrent event because some outcomes cease to exist after death.

Suppose the variable is a physical-function score at Month 12.

A patient who dies at Month 8 cannot have a Month 12 physical-function measurement.

Possible estimand strategies include:

• Composite: death is considered a poor outcome.
• Hypothetical: estimate the Month 12 score under a specified hypothetical survival scenario.
• Other strategies depending on the clinical question and endpoint.

Simply labeling the Month 12 score as “missing” does not resolve the clinical meaning of death.

Estimands for Time-to-Event Endpoints

Time-to-event endpoints already incorporate certain events into the endpoint definition.

For example, progression-free survival may define an event as disease progression or death.

In such cases, the endpoint itself can represent a composite strategy for the relevant events, depending on the precise estimand specification.

However, treatment discontinuation, rescue medication, treatment switching, or other post-randomization events may still require explicit consideration.

Estimand and Censoring

Censoring should not be used as an automatic substitute for an intercurrent-event strategy.

For example, censoring patients at treatment discontinuation changes which observations contribute to the estimator and may correspond to a while-on-treatment question rather than a treatment-policy question.

The analyst should therefore ask:

If the implied question is different from the prespecified estimand, the analysis is misaligned.

Practical Estimand Specification Table

A highly effective implementation technique is to create an estimand table before writing the detailed analysis methods.

Estimand Matrix for Multiple Intercurrent Events

A trial can have several intercurrent events, and each may require a different strategy.

The matrix should be finalized before the primary analysis is programmed.

From Estimand to SAP

The SAP should translate the estimand into operational language.

A useful structure is:

Example SAP Language

This wording is illustrative. The exact model, covariates, covariance structure, missing-data assumptions, estimand population, and event rules must be tailored to the protocol and SAP.

Estimand and ADaM Implementation

For a CDISC-oriented workflow, the estimand should be traceable from the clinical specification into the analysis dataset and output.

A useful traceability chain is:

Useful Analysis Variables

The exact variable names depend on the study, but an analysis dataset may need variables such as:

The names above are examples only. Production datasets should follow the study's approved ADaM specification and naming conventions.

R Implementation: Treatment-Policy Example

Suppose a longitudinal dataset contains one row per patient per analysis visit.

library(dplyr)
library(lme4)
library(emmeans)

analysis_data <- adlb %>%
  filter(ANL01FL == "Y") %>%
  mutate(
    TRT01P = factor(TRT01P),
    AVISIT = factor(AVISIT),
    BASE = as.numeric(BASE),
    AVAL = as.numeric(AVAL),
    CHG = AVAL - BASE
  )

fit <- lmer(
  CHG ~ BASE + TRT01P * AVISIT + (1 | USUBJID),
  data = analysis_data
)

emm <- emmeans(
  fit,
  ~ TRT01P | AVISIT
)

contrast(
  emm,
  method = "revpairwise",
  at = list(AVISIT = "Week 24")
)

The code illustrates an analysis workflow, not a universal recommendation. The model must be chosen to target the prespecified estimand and should follow the approved SAP.

Why the R Code Alone Does Not Define the Estimand

Consider the same model applied to two different datasets:

• Dataset A includes post-rescue outcomes because the estimand is treatment policy.
• Dataset B excludes post-rescue outcomes because the estimand is while on treatment.

The software procedure could be identical while the estimands are different.

Therefore, the estimand is defined by the scientific and operational specification, not by the function call.

R Implementation: Estimand Metadata

For reproducibility, an analysis project can store the estimand specification as metadata.

estimand <- list(
  population = "All randomized eligible patients",
  treatment = "Experimental vs placebo",
  variable = "Change from baseline at Week 24",
  intercurrent_events = list(
    treatment_discontinuation = "Treatment policy",
    rescue_medication = "Treatment policy",
    death = "Composite"
  ),
  summary = "Difference in means",
  primary_estimator = "Prespecified repeated-measures model"
)

estimand

This simple pattern can make the relationship between the clinical specification and the programming implementation more explicit.

Example: Creating an Intercurrent-Event Flag

An analysis dataset may retain the event date and a derived indicator rather than deleting post-event observations automatically.

analysis_data <- analysis_data %>%
  mutate(
    after_rescue =
      !is.na(RESCDT) & ADT > RESCDT,

    after_discontinuation =
      !is.na(DCDT) & ADT > DCDT
  )

Under a treatment-policy strategy, these flags can be used for descriptive checks while the relevant post-event outcomes remain eligible for the primary analysis, subject to the SAP.

Example: While-on-Treatment Rule

A while-on-treatment analysis may define an analysis window relative to treatment exposure.

analysis_data <- analysis_data %>%
  mutate(
    within_exposure =
      ADT <= DCDT
  ) %>%
  filter(
    is.na(DCDT) | within_exposure
  )

This is intentionally simplified. Real trials often require grace periods, exposure windows, rescue rules, partial dates, and protocol-specific definitions.

Example: Composite Outcome

For a binary response endpoint, a composite rule might be programmed as:

analysis_data <- analysis_data %>%
  mutate(
    RESP_COMP = case_when(
      DEATHFL == "Y" ~ 0,
      RESPONSE == "Y" & ALIVE_WK24 == "Y" ~ 1,
      TRUE ~ 0
    )
  )

The important point is that the composite definition is part of the estimand. It should be specified clinically before the result is examined.

Sensitivity Analysis in R

A sensitivity analysis should change a meaningful assumption, not merely produce a different p-value.

For example, a simple analysis could compare estimates under different analysis datasets or missing-data assumptions:

primary_fit <- lmer(
  CHG ~ BASE + TRT01P * AVISIT + (1 | USUBJID),
  data = primary_data
)

sensitivity_fit <- lmer(
  CHG ~ BASE + TRT01P * AVISIT + (1 | USUBJID),
  data = sensitivity_data
)

emmeans(primary_fit, ~ TRT01P | AVISIT)
emmeans(sensitivity_fit, ~ TRT01P | AVISIT)

In a regulated analysis, the sensitivity analysis definitions should be prespecified and documented rather than created after reviewing the primary result.

Common Estimand Implementation Errors

1. Starting with the statistical model. The model should follow the estimand, not define it.
2. Using “ITT” as if it completely specifies the estimand. An intention-to-treat population or principle does not by itself resolve every intercurrent event.
3. Treating treatment discontinuation as automatically missing data. The clinical meaning of discontinuation must first be addressed.
4. Deleting post-rescue observations without stating the strategy. This may change the treatment question.
5. Calling an MMRM the estimand. MMRM is an estimator/modeling approach.
6. Using “hypothetical” without defining the hypothetical scenario. The counterfactual condition must be scientifically meaningful.
7. Calling a complete-case analysis treatment policy. Excluding patients with missing outcomes can change the target population or require assumptions.
8. Changing the estimand after looking at the data. The treatment-effect question should be defined prospectively.
9. Failing to distinguish estimand from endpoint. An endpoint alone does not fully define the treatment effect.
10. Providing sensitivity analyses unrelated to the primary assumptions. Sensitivity analyses should probe material uncertainty in the inference.

Estimand Alignment Checklist

Estimand Traceability Matrix

A practical governance tool is a traceability matrix connecting the estimand to downstream deliverables.

How to Review an Existing SAP for E9(R1) Alignment

When reviewing an existing SAP, search for the following concepts:

• Treatment discontinuation
• Rescue medication
• Concomitant medication
• Treatment switching
• Death
• Missing data
• Censoring
• Post-treatment assessments
• Intercurrent event
• Hypothetical scenario
• Treatment-policy language

Then ask whether each rule is answering a clinical question or merely describing a programming operation.

A Practical Review Example

Suppose an SAP states:

That sentence is incomplete from an E9(R1) perspective.

The reviewer should ask:

1. What is the clinical question after discontinuation?
2. Is the intended strategy while on treatment?
3. Is discontinuation being treated as an intercurrent event?
4. Are outcomes after discontinuation irrelevant, or merely missing?
5. What population-level treatment effect is being estimated?
6. What assumptions justify the estimator?

Only after answering those questions should the data-exclusion rule be finalized.

When Multiple Estimands Are Appropriate

A clinical trial may legitimately have more than one estimand.

For example:

The important requirement is that each estimand is explicitly defined and that the primary and supportive analyses are labeled according to the questions they answer.

Estimands and Subgroup Analyses

A subgroup analysis can target an estimand within a prespecified population defined by baseline characteristics.

For example, an estimand might be defined separately for patients with and without a baseline biomarker.

However, subgroup membership should generally be based on variables defined before randomization or according to the prespecified scientific framework. Post-randomization subgroup definitions can introduce additional causal complications.

Estimands and Multiplicity

If multiple estimands are part of confirmatory decision-making, the statistical strategy should explain which estimand is primary and how multiplicity is handled when applicable.

Not every descriptive estimand requires multiplicity adjustment, but confirmatory claims should follow the prespecified inferential framework.

Estimands and Clinical Interpretation

The final interpretation should use language that matches the estimand.

For example, avoid reporting:

when the analysis actually estimates a treatment-policy effect, unless the context makes the treatment-policy interpretation clear.

A stronger interpretation is:

What Makes an Estimand “Good”?

A useful estimand should be:

• Clinically meaningful: it addresses a real decision or scientific question.
• Precise: the five attributes are sufficiently specific to avoid ambiguity.
• Estimable: the trial collects information needed to support estimation.
• Interpretable: the resulting treatment effect has a clear clinical meaning.
• Aligned: design, conduct, analysis, and interpretation all target the same question.

Practical Implementation Workflow

A Compact Worked Scenario

Consider a randomized trial of Treatment A versus placebo with a continuous outcome at Week 24. Rescue medication is allowed from Week 12 onward.

The clinical team asks:

The estimand is naturally aligned with a treatment-policy strategy.

The implementation consequences are:

• Do not automatically remove patients after treatment discontinuation.
• Continue collecting outcomes after rescue when feasible.
• Define how death affects the endpoint if the outcome becomes undefined.
• Use an estimator appropriate for the resulting longitudinal data and missingness assumptions.
• Conduct sensitivity analyses for missing outcomes and other material assumptions.

This example illustrates the central E9(R1) idea: the estimand changes what data the trial needs and how the analysis should be designed.

Validation and Quality Control

Estimand implementation should be validated at more than one level.

Estimand QC Questions for Programmers

• Which patients are included?
• Which treatment variable is used?
• Which observations remain eligible after treatment discontinuation?
• Which observations remain eligible after rescue?
• How is death represented?
• Are post-event measurements retained or excluded?
• What is the analysis visit definition?
• Which covariates are prespecified?
• What treatment contrast is extracted from the model?
• Does the output footnote use the same estimand language as the SAP?

Estimand QC Questions for Statisticians

• Is the clinical question precise?
• Are all important intercurrent events identified?
• Is each event assigned an explicit strategy?
• Does the chosen estimator target the estimand?
• Are the assumptions plausible and documented?
• Are sensitivity analyses informative?
• Is the population-level summary clinically interpretable?

Estimand QC Questions for Clinical Teams

• Would the resulting treatment effect answer the question used for treatment decisions?
• Does the treatment-policy versus hypothetical distinction make clinical sense?
• Should rescue medication count as treatment failure?
• Should death be incorporated into the outcome?
• Are post-discontinuation outcomes clinically meaningful?
• Would the interpretation be understandable to clinicians and patients?

References and Official Guidance

International Council for Harmonisation. E9(R1) Addendum on Estimands and Sensitivity Analysis in Clinical Trials to the Guideline on Statistical Principles for Clinical Trials. Final version, adopted 20 November 2019.
Official ICH E9(R1) guideline

U.S. Food and Drug Administration. E9(R1) Statistical Principles for Clinical Trials: Addendum: Estimands and Sensitivity Analysis in Clinical Trials. Guidance for Industry, May 2021.
FDA guidance page

U.S. Food and Drug Administration. E9 Statistical Principles for Clinical Trials. September 1998.