ICH E9(R1) Estimand Framework
Definition
The estimand framework was introduced in ICH E9(R1) (2019) to align trial planning, design, conduct, analysis, and interpretation. It provides a precise description of what will be estimated and how.
ICH E9(R1) Definition:
"An estimand is a precise description of the treatment effect reflecting the clinical question posed by a given trial objective. It includes specification of which outcome is of interest, in which patients, over what period, and how it will be summarized."
The estimand framework fundamentally shifts oncology trial thinking from "what data do we collect?" to "what clinical question do we answer?" This requires explicit specification before trial initiation, not post-hoc.
The Four Estimand Attributes
Every estimand must specify four attributes:
1. Population (Target Population)
Definition: Which patients' outcomes will be included in the treatment effect estimate.
| Specification | Description | Oncology Example |
|---|---|---|
| Intent-to-treat (ITT) | All randomized patients, regardless of adherence | Standard for regulatory approval |
| Modified ITT (mITT) | All randomized patients minus major protocol violators | Justified pre-specification required |
| Per-protocol (PP) | Patients who adhered to protocol | Sensitive to selection bias; secondary only |
| Principal stratum | Patients who would not have experienced the IE | Effect in "tolerant" or "compliant" subgroup |
| Biomarker-positive | Pre-specified biomarker+ subgroup | PD-L1+, MSI-H, mutation-specific therapies |
Regulatory Preference: ITT is FDA's preference because it reflects real-world heterogeneity and avoids selection bias from post-randomization exclusions.
SAP Language Example:
"The primary analysis population is the intent-to-treat (ITT) population, defined as all patients randomized, analyzed as randomized regardless of treatment adherence, discontinuation, or protocol deviation."
2. Variable (Endpoint / Outcome of Interest)
Definition: What is measured for each patient.
| Variable Type | Definition | Oncology Examples |
|---|---|---|
| Time-to-event | Time from a defined origin to occurrence of event | OS, PFS, DFS, TTR |
| Continuous | Numerical measurement at fixed time points | Biomarker change, PRO score, blood counts |
| Binary | Presence/absence of outcome | ORR, CR, safety event |
| Composite | Two or more outcomes combined into single variable | Death or progression, DFS (relapse + death) |
Specification Required:
- Origin point: Randomization, treatment start, or landmark date
- Event definition: What constitutes occurrence of outcome
- Assessment schedule: When and how often measured
- Primary endpoint vs. secondary endpoints
SAP Language Example:
"The primary efficacy endpoint is Overall Survival (OS), defined as time from randomization to death from any cause. Patients alive at final data cutoff will be censored at their last known alive date."
3. Intercurrent Events and Handling Strategies
Definition: How post-randomization events that affect interpretation or existence of the outcome are addressed.
See Intercurrent Events in Oncology Trials for comprehensive IE strategy details. Briefly:
| Strategy | What It Does | When Used |
|---|---|---|
| Treatment Policy | Measure outcome regardless of IE; pragmatic real-world effect | Primary for most oncology trials; FDA preference |
| Hypothetical | Estimate outcome as if IE had not occurred (counterfactual) | Subsequent therapy, crossover (OS adjustment) |
| Composite | Redefine outcome to include the IE | Death before progression; terminal IE |
| Principal Stratum | Estimate effect in subpopulation where IE would not occur | Tolerability in adherent patients (rarely used) |
| While-on-Treatment | Measure outcome only during treatment period | Symptom-based, QoL endpoints |
Critical Distinction: IE strategy determines the clinical question being answered. Different IEs in the same trial may require different strategies.
Example — OS Estimand with Subsequent Therapy IE:
- Primary: Treatment Policy (OS measured regardless of subsequent therapy)
- Sensitivity: Hypothetical with RPSFT adjustment (OS if subsequent therapy had not occurred)
SAP Language Example:
"Subsequent anticancer therapy is an intercurrent event. Per the treatment policy strategy, OS will be measured and reported regardless of whether patients receive subsequent therapy post-progression. All randomized patients will be included in the primary OS analysis."
4. Population-Level Summary (Analysis Summary)
Definition: How individual patient outcomes are combined to produce a single treatment effect estimate.
| Summary Type | Description | Oncology Use | Interpretation |
|---|---|---|---|
| Hazard Ratio (HR) | Relative instantaneous rate of event occurrence | OS, PFS (time-to-event) | HR < 1 = treatment benefit |
| Restricted Mean Survival Time (RMST) | Mean survival within pre-specified time window | OS, PFS (robust to non-PH) | Difference = additional months of survival |
| Median Survival | Time at which 50% have experienced event | OS, PFS (descriptive) | Reported with confidence intervals |
| Difference in Proportions | Treatment group % event - Control group % event | ORR, safety (binary) | Difference = absolute benefit |
| Odds Ratio | Odds of event in treatment vs. control | ORR, CR (binary) | OR > 1 = treatment benefit |
| Event Rates | Landmark rates at pre-specified times | 1-year OS, 2-year PFS | Reported per arm with CIs |
Statistical Methods Implied:
- Time-to-event: Kaplan-Meier curves, log-rank test, Cox regression
- Continuous: t-test, ANCOVA, mixed model for repeated measures
- Binary: Chi-square test, logistic regression
- Composite: Competing risks (Fine-Gray), cumulative incidence
SAP Language Example:
"The primary efficacy analysis of OS will use the intent-to-treat population. Kaplan-Meier curves will be stratified by [randomization factor]. A stratified log-rank test will be used for the primary hypothesis test. Hazard ratios and 95% confidence intervals will be reported from a Cox proportional hazards model, stratified by [factor]."
The Five Intercurrent Event Strategies in Detail
For comprehensive strategy-by-strategy guidance, see Intercurrent Events in Oncology Trials. Summary below:
1. Treatment Policy Strategy
When Dominant: OS primary endpoint; treatment discontinuation due to adverse events
Rationale: Reflects real-world clinical use and pragmatic effect of assigned treatment
Statistical Implementation:
- No censoring related to the IE; standard outcome censoring applies
- Full ITT population
- Data collection continues post-IE
- Kaplan-Meier, Cox regression, or other standard methods
2. Hypothetical Strategy
When Dominant: Subsequent therapy confounds OS; crossover in open-label trials
Rationale: Isolates intrinsic treatment effect absent the confounding IE
Statistical Implementation:
- Censor at IE occurrence
- For crossover: RPSFT, 2SRST, or IPCW adjustment methods
- Requires strong structural assumptions (AFT model, exchangeability)
- Typically sensitivity, not primary
3. Composite Strategy
When Dominant: Terminal IE (death before progression); severe safety event
Rationale: IE precludes outcome measurement; redefine to include IE as component
Statistical Implementation:
- Redefine outcome to include both original event and IE
- Competing risks analysis (Fine-Gray) for sensitivity
- ITT population
- Kaplan-Meier or cumulative incidence curves
4. Principal Stratum Strategy
When Dominant: Efficacy in drug-tolerant subgroup; adherence analysis
Rationale: Clinical interest is effect in patients who would comply with treatment
Statistical Implementation:
- Define stratum pre-specification (baseline or early post-baseline)
- Subgroup analysis or causal inference methods
- Risk of selection bias; strong sensitivity analyses required
- Rarely used in oncology; weak regulatory precedent
5. While-on-Treatment Strategy
When Dominant: Symptom-based endpoints; QoL during treatment period
Rationale: Clinical question is effect while patient receives treatment
Statistical Implementation:
- Restrict follow-up to time on treatment
- Censor at treatment discontinuation
- Report time-to-treatment-discontinuation distribution
- Risk of informative censoring
Integration of Estimands into Phase 3 Statistical Analysis Plans
Successful Phase 3 trials require explicit estimand specification in both protocol and Statistical Analysis Plan (SAP), with agreement from regulators before trial launch.
Phase 3 Protocol Requirements
ICH E9(R1) §6.2.3 — Protocol and SAP:
"The protocol should specify the primary estimand(s) of interest. The analysis plan should specify the main estimator aligned with each estimand."
Minimum Protocol Content:
-
Primary Estimand Specification (Section 8, Statistical Methods):
- Population: ITT vs. mITT vs. PP (justify choice)
- Variable: Primary endpoint definition with assessment schedule
- Intercurrent Events:
- List all anticipated IEs
- Pre-specify handling strategy for each
- Justify strategy choice per clinical question
- Population-level Summary: Statistical method (HR, RMST, etc.)
-
Secondary and Supportive Estimands:
- Alternative strategies for sensitivity (e.g., hypothetical for treatment policy primary)
- Subgroup estimands if pre-specified
- Safety endpoints with IE strategy
-
Sensitivity Analysis Plan (Section 10.2):
- List all pre-specified sensitivity analyses
- One assumption per sensitivity (vary one at a time)
- Tipping point analysis criteria (when applicable)
- Reference Sensitivity Analyses for Estimands for hierarchy
Example Protocol Language:
Primary Estimand — Overall Survival: - Population: Intent-to-treat (ITT) population - Variable: Overall Survival (OS) — time from randomization to death from any cause - Intercurrent Events: - Subsequent anticancer therapy is anticipated in ~40% of control patients - Strategy: Treatment Policy — OS will be measured and reported regardless of subsequent therapy received - Rationale: Reflects real-world efficacy and aligns with FDA guidance for OS as safety-related endpoint - Population-Level Summary: Hazard ratio from Cox model, stratified by [prognostic factor]; Kaplan-Meier curves reported
Sensitivity Estimand — Hypothetical OS (without subsequent therapy): - Population: Same as primary (ITT) - Variable: OS, hypothetical scenario where subsequent therapy does not confound - IE Handling: Hypothetical strategy with RPSFT adjustment - Rationale: Robustness check; explores mechanistic effect absent confounding - Analysis Population-Level: RPSFT-adjusted HR with 95% CI; presented as supportive
Statistical Analysis Plan (SAP) Requirements
The SAP must be finalized and locked BEFORE DATABASE LOCK (ideally before unblinding in blinded trials).
SAP Sections Specific to Estimands:
Section A: Estimand Specifications (Before Section 7)
Create a table for each primary, secondary, and sensitivity estimand:
| Estimand Component | Specification |
|---|---|
| Estimand Name | OS (Treatment Policy) |
| Population | ITT (all randomized) |
| Endpoint | Time to death from any cause |
| Intercurrent Event 1 | Subsequent therapy; Strategy: Treatment Policy |
| Intercurrent Event 2 | Treatment discontinuation; Strategy: Treatment Policy |
| Summary Measure | Hazard ratio (Cox model, stratified by site) |
| Primary Estimator | Stratified log-rank test; Cox regression |
| Censoring Rule | Administratively censored at LKDA if event not observed |
| Missing Data Assumption | MCAR; standard methods apply |
| Sensitivity Analyses | RPSFT for subsequent therapy IE; RMST at 24 months |
Section 7.2: Primary Efficacy Analysis
Standard SAP Language:
7.2.1 Primary Estimand and Analysis
The primary efficacy analysis will use the intent-to-treat (ITT) population under the treatment policy strategy for intercurrent events.
Endpoint: Overall Survival (OS) is defined as the time from randomization to death from any cause. Patients alive at the final data cutoff will be censored at the date of last known alive.
Analysis Population: All randomized patients (N=[X]), analyzed as randomized.
Intercurrent Event Handling: - Subsequent anticancer therapy: Treatment Policy (OS measured regardless) - Treatment discontinuation: Treatment Policy (continued follow-up) - Missing OS status: Handled per [protocol Section X.X]
Primary Analysis: 1. Kaplan-Meier curves stratified by [randomization factor] 2. Stratified log-rank test (primary hypothesis test) 3. Cox proportional hazards model with stratification; report HR (95% CI) 4. Median OS and 1-year OS rates per arm with 95% CIs
Significance Level: Two-sided α = 0.05 (not adjusted for multiplicity of sensitivity analyses)
Section 8: Sensitivity Analyses
8.1 Sensitivity Analyses for OS (Treatment Policy)
The following pre-specified sensitivity analyses will be conducted to evaluate robustness of the primary OS estimate:
Sensitivity 1: Alternative Censoring Rule - Censor at last contact date (vs. LKDA in primary) - Rationale: Tests robustness to censoring definition
Sensitivity 2: RMST Analysis (Non-Proportional Hazards Robustness) - Restricted mean survival time at 24 months - Rationale: Does not assume proportional hazards; more robust if treatment effect emerges late
Sensitivity 3: RPSFT Adjustment for Subsequent Therapy - Rank-Preserving Structural Failure Time adjustment - Rationale: Tests hypothetical OS scenario absent subsequent therapy confounding - Method: G-estimation; Bootstrapped 95% CIs (1,000 iterations) - Assumptions: AFT model holds; randomization preserved; no unmeasured confounding
Sensitivity 4: Landmark Analysis - OS rates at 12, 24, 36 months per arm - Rationale: Identifies potential non-proportional hazards or differential follow-up
Sensitivity 5: Subgroup Consistency - Treatment effect in pre-specified subgroups (ECOG PS, age, prior lines) - Rationale: Confirms consistency of benefit across population
Indication-Specific Estimand Patterns (2020–2025)
Regulatory practice has converged on indication-specific estimand choices:
NSCLC (Metastatic and Adjuvant)
OS Estimand:
- Primary: Treatment Policy (subsequent therapy is standard-of-care)
- Sensitivity: Hypothetical with RPSFT if crossover > 20%
PFS Estimand:
- Primary: Treatment Policy (discontinuation and dose modification captured)
- Alternative: Hypothetical if subsequent therapy is key confounding question
Breast Cancer
OS Estimand:
- Primary: Treatment Policy (heterogeneous post-progression therapies)
- PFS Estimand: Treatment Policy (standard)
Adjuvant DFS:
- Primary: Composite (relapse or death); ITT population
Hematologic Malignancies (AML, Lymphoma, Myeloma)
OS Estimand:
- Primary: Treatment Policy (HSCT, CAR-T access standard; FDA accepts despite transformation)
- Rationale: HSCT eligibility reflects disease biology and patient factors; treatment policy captures pragmatic effect
Event-Free Survival or Progression-Free Survival:
- Primary: Composite (relapse/progression or death)
Rare Diseases and Pediatric Oncology
OS Estimand:
- Primary: Treatment Policy (small sample sizes; every patient matters)
- Sensitivity: Per-protocol analysis if major protocol violations occur
Regulatory Expectations and FDA/EMA Positions
FDA Position (Post-2021 Adoption of ICH E9(R1))
Mandatory Requirements:
- Pre-specification: Estimand in protocol and SAP before database lock
- Treatment Policy Default: Unless strong clinical justification, OS should use treatment policy
- Sensitivity Analyses: Required for all primary estimands; at minimum, one alternative strategy
- Transparency: All assumptions stated explicitly and justified
OS as Safety Endpoint:
- FDA's 2025 OS draft guidance reinforces treatment policy as primary for OS
- RMST or weighted log-rank required if non-proportional hazards suspected
Controversial Choices (Risk of Rejection):
- Hypothetical OS as primary without RPSFT or structural method
- Principal stratum without strong clinical rationale
- Missing sensitivity analyses for high-missing-data scenarios
- Post-hoc estimand specification
EMA Position
Similar to FDA:
- Clear estimand specification mandatory pre-submission
- Treatment policy preferred for OS and PFS
- Hypothetical acceptable if mechanistic question is strong
Differences:
- More accepting of hypothetical strategy for targeted therapies (mechanistic question)
- Emphasis on sensitivity hierarchy over single primary
PMDA (Japan) Position
- Adopted ICH E9(R1) in 2019
- Estimand specification required in Japanese CTD dossier
- Alignment with FDA/EMA on treatment policy default
SAP Checklist for Estimand Compliance
Use this checklist when finalizing Phase 3 SAP:
- [ ] Primary estimand fully specified: Population, variable, IE strategies, summary measure
- [ ] Secondary estimands listed: At least one alternative estimand or strategy
- [ ] Sensitivity analyses pre-specified: One assumption per sensitivity (min. 3–5 analyses)
- [ ] Tipping point criteria defined: If missing data burden > 10% or OS borderline significant
- [ ] IE strategy justified: Rationale for choice relative to clinical question
- [ ] Multiplicity statement: Sensitivity analyses do not contribute to Type I error control
- [ ] Missing data assumptions stated: MAR, MNAR, imputation method, tipping point approach
- [ ] Analysis population alignment: ITT, mITT, or PP clearly defined and justified
- [ ] Statistical methods aligned: Method (Cox, RMST, etc.) matches estimand summary measure
- [ ] Regulatory reference: Cite FDA/EMA/ICH E9(R1) section justifying approach
- [ ] Protocol consistency: SAP matches protocol estimand section; no post-hoc changes
Backlinks
Oncology Endpoint Overview Intercurrent Events in Oncology Trials Principal Stratum and While-on-Treatment Strategies Sensitivity Analyses for Estimands Overall Survival (OS) Progression-Free Survival (PFS)
Source: ICH Harmonised Guideline E9(R1) — Addendum on Estimands and Sensitivity Analysis in Clinical Trials (Final, November 2019) Status: Final (ICH E9(R1) Step 4, November 2019; FDA adoption May 2021) Compiled from ICH E9(R1) §1–6 + FDA regulatory feedback + implementation patterns (2020–2025)
Last Updated: April 2026
Knowledge Base: oncology_kb
Section: Clinical Trial Design — Estimand Framework and Phase 3 Integration