Intercurrent Events in Oncology Trials

Definition

An intercurrent event (IE) is a post-randomization event that affects interpretation or the existence of the outcome of interest in a clinical trial. In oncology, IEs are ubiquitous and fundamentally shape trial design and estimand specification.

ICH E9(R1) Definition (Final, 2019):

"An intercurrent event is a post-randomization event (either observed or unobserved) that either precludes the collection of the variable of interest or affects its interpretation."

Common oncology IEs include:

Treatment discontinuation (due to adverse event, disease progression, or patient choice)
Subsequent anticancer therapy (confounding OS measurement)
Dose reduction or modification (PFS definition impact)
Crossover (open-label trials)
Death before progression (competing risk)
Rescue or supportive medication (endpoint contamination)
Protocol deviation (noncompliance, eligibility violations)
External events (COVID, supply chain disruption)

The choice of IE handling strategy determines what clinical question is answered and which analysis population applies.

The Estimand Framework Context

ICH E9(R1) defines an estimand as:

"a precise description of the treatment effect reflecting the clinical question posed by a given trial objective."

An estimand has four attributes:

Attribute	Definition	Oncology Example
Population	Patients targeted by the clinical question	ITT population, biomarker-positive subgroup, principal stratum
Variable (endpoint)	What is measured for each patient	PFS, OS, ORR, PRO score
Intercurrent events and handling strategies	How IEs are incorporated into the estimand	Treatment policy for discontinuation; hypothetical for crossover
Population-level summary	How individual outcomes are combined	Hazard ratio, difference in proportions, restricted mean survival time

The choice of IE strategy directly shapes what clinical question is being answered. Different IEs in the same trial may warrant different strategies.

IE Categories by Functional Impact

Two functional categories of IEs follow from the ICH E9(R1) definition:

Category	What the IE does	Example
Affects interpretation	Outcome exists but its meaning is confounded	Subsequent anticancer therapy commenced; patient takes rescue medication
Affects existence	Outcome cannot be measured or is undefined	Death before a PRO assessment; withdrawal from study before progression

Critical Distinction: IE vs. Missing Data

A distinction emphasized by ICH E9(R1): discontinuation of randomised treatment is an intercurrent event (addressed by strategy choice), whereas withdrawal from the study gives rise to missing data (addressed by imputation or informative censoring methods). These are conceptually and analytically separate.

The Five Intercurrent Event Strategies

The ICH E9(R1) framework specifies five distinct strategies for handling IEs, each answering a different clinical question:

1. Treatment Policy Strategy

Definition (ICH E9(R1)):

The treatment policy strategy stipulates that intercurrent events are considered irrelevant to the treatment effect definition: the outcome is measured and reported regardless of whether the intercurrent event occurs.

When to Use in Oncology:

Primary endpoint for regulatory approval (FDA preference in most settings)
Treatment discontinuation due to adverse events (measures real-world effect)
Dose modifications when ongoing measurement is feasible
COVID or other external interruptions (sensitivity to trial disruption)
When ITT principle is clinically and regulatory favored

Statistical Implementation:

Censoring: No censoring related to the IE itself; standard progression/death censoring applies
Analysis population: Full ITT (or modified ITT)
Follow-up: Data collection continues regardless of IE occurrence
SAP language: "Outcomes will be measured and reported regardless of treatment discontinuation or modification. Patients will be followed per protocol."

Clinical Rationale:

Reflects the pragmatic effect of the assigned treatment in real-world settings where IEs occur. Particularly relevant in oncology where tolerability and adherence are key considerations for approval.

2. Hypothetical Strategy

Definition (ICH E9(R1)):

The hypothetical strategy defines the treatment effect under the hypothetical scenario that the intercurrent event would not occur. The objective is to estimate what the outcome would have been if the intercurrent event had not happened (a counterfactual).

When to Use in Oncology:

Subsequent anticancer therapy (estimates effect without confounding post-progression therapy)
Rescue medication (estimates intrinsic drug effect)
Crossover in open-label trials (OS adjusted via RPSFT or 2SRST)
Protocol deviations (sensitivity to protocol adherence)
Mechanism-of-action questions (does drug work if patient stays on it?)

Statistical Implementation:

Censoring: Censor at IE occurrence (e.g., censoring OS at second-line therapy start)
Analysis population: Patients without IE (or at IE in sensitivity)
Imputation/adjustment: RPSFT, 2SRST, or IPCW for structural missing data; rank-preserving or inverse probability weighting
Assumptions required: Strong structural or causal assumptions (e.g., AFT model for RPSFT)

3. Composite Strategy

Definition (ICH E9(R1)):

The composite strategy redefines the variable of interest such that the intercurrent event becomes a component of the endpoint itself. This is commonly used when the IE is a terminal event (e.g., death) that precludes measurement of the original outcome.

When to Use in Oncology:

Death before progression (composite endpoint: death or progression)
Severe adverse events requiring discontinuation (composite: disease progression or unacceptable toxicity)
Rescue therapy in symptom-driven trials (composite: symptom progression or need for rescue)
Combined efficacy–safety assessment (death, progression, or major AE)

Statistical Implementation:

New outcome definition: Combine the IE occurrence with the original outcome (e.g., "death or progression")
Event hierarchy: Specify priority if IE and outcome occur near-simultaneously (e.g., if death within 7 days of progression, count as death)
Analysis population: Full ITT
Competing risks: Account for competing events (e.g., death from non-cancer causes competes with progression)

Example: Death Before Progression in Phase 3 Metastatic Setting

Original outcome: Progression-free survival (PFS)
IE: Death from any cause
Composite strategy: PFS2 or death-or-progression (DFS), where event = death OR radiologically confirmed progression (whichever occurs first)
Interpretation: Reflects both efficacy (delay of progression) and tolerability (avoidance of early death)

4. Principal Stratum Strategy

Definition (ICH E9(R1)):

The principal stratum strategy redefines the population of interest such that the intercurrent event is a characteristic of the principal stratum being studied, and it is of clinical interest to estimate the treatment effect in the subpopulation where the IE would not occur.

When to Use in Oncology:

Tolerability studies (effect in patients who can tolerate full dose)
Patients able to complete planned therapy (adherent subgroup)
ECOG 0–1 subset (performance status responders)
Biomarker-positive subgroup (for targeted therapies)
Safety analyses (effect in patients without treatment-limiting AEs)

Statistical Implementation:

Subgroup definition: Specify baseline or early-post-baseline criteria that define the principal stratum (e.g., "patients with no Grade ≥3 AE by Week 4")
Caution—Selection bias: Principal stratification can introduce bias if the stratum is defined post-randomization and is related to treatment
Assumptions: Strong causal assumptions; often requires untestable assumptions about consistency and no unmeasured confounding within stratum

Example: Efficacy in Treatment-Tolerant Patients

Question: What is the treatment effect in patients who tolerate full-dose therapy for ≥80% of planned duration?
Principal stratum: Patients who remain on treatment
Statistical approach: Subgroup analysis (if IE is baseline-defined) or causal inference methods (if post-baseline); sensitivity analyses for selection bias

5. While-on-Treatment Strategy

Definition (ICH E9(R1)):

The while-on-treatment strategy defines the treatment effect during the period while the intercurrent event has not yet occurred. Measurements are collected only while the IE is absent.

When to Use in Oncology:

Symptom-based endpoints (while on study drug only)
Quality of life (while patient is receiving treatment)
Dose intensity or exposure (safety surrogate)
Biomarker measurements during treatment period

Statistical Implementation:

Follow-up period: Restrict analysis to time on treatment
Censoring: Censor at treatment discontinuation
Survival curves: Kaplan-Meier curves will show rapid rise in censoring after IE
SAP note: Must report time-to-IE distribution to contextualize results

Common IEs in Oncology — Reference Table

Intercurrent Event	Frequency in Oncology	Common Strategy(ies)	Statistical Consequence	Measurement Challenge
Tx discontinuation (AE)	Very common (10–40%)	Treatment Policy	Defines ITT follow-up	Continue or stop measuring?
Subsequent anticancer Tx	Very common (30–70% OS trials)	Treatment Policy (primary); Hypothetical (sensitivity, RPSFT/IPCW)	OS confounding, requires adjustment	Blinding loss, confounding
Dose reduction/modification	Common (20–50%)	Treatment Policy	PFS definition impact, exposure-response	PFS measured as-treated
Death before progression	Common (5–15% metastatic)	Composite	Competing risk, endpoint power	Timing of death vs. radiological assessment
Crossover (open-label)	Oncology-specific (50–80% crossover trials)	Hypothetical (RPSFT/2SRST)	OS dilution, strong assumptions	Blinding not possible, bias risk
Rescue/supportive medication	Variable (symptom-driven trials)	Composite or Hypothetical	Symptom suppression, unmeasured confounding	Masking of true effect
Protocol deviation	Inherent (ITT vs. per-protocol)	Treatment Policy (ITT, primary)	Effect estimation; per-protocol may introduce bias	Selection bias in per-protocol
External event (COVID, supply chain)	Situational (2020+)	Sensitivity (exclude period or hypothetical)	Trial integrity, generalizability	Uncontrollable variable

Practical Decision Framework: Which Strategy for Which IE?

Strategy Selection by IE Type

Intercurrent Event	Regulatory Preference	Primary Strategy	Sensitivity Strategy	Key Assumption
Tx discontinuation (AE)	Treatment Policy (ITT)	Measure outcomes post-DC	Hypothetical (censor at DC)	Real-world tolerability matters
Subsequent anticancer Tx	Treatment Policy or Hypothetical	Measure OS regardless	Hypothetical (RPSFT/2SRST/IPCW)	Subsequent therapy confounds OS
Dose reduction	Treatment Policy	Continue assessment at reduced dose	Principal stratum (full-dose cohort)	Dose modification is part of strategy
Death before progression	Composite	DFS (death or progression)	None needed (IE is outcome)	Death prevents measuring other outcomes
Crossover (open-label)	Hypothetical (investigator-blinded)	RPSFT or 2SRST adjustment	Treatment policy (unadjusted)	AFT holds or 2nd randomization model
Rescue/supportive Tx	Composite or Hypothetical	Composite (outcome + rescue)	Hypothetical (censor at rescue)	Rescue masks true benefit or harm
Protocol deviation	Treatment Policy (ITT)	Full ITT population	Per-protocol (principal stratum)	Protocol adherence drives effect?
COVID or external event	Sensitivity analysis	Treatment Policy (primary)	Hypothetical (exclude period)	Event was uncontrollable/non-clinical

Decision Algorithm

Is the IE terminal? (e.g., death)
→ Use Composite (death becomes part of endpoint) or Treatment Policy (stop measuring)
Does the IE confound the outcome? (e.g., subsequent therapy for OS)
→ Use Hypothetical (adjust via RPSFT/2SRST/IPCW) with Treatment Policy as sensitivity
Is the IE clinically integral to the treatment strategy? (e.g., dose reduction, rescue therapy)
→ Use Treatment Policy (primary) with Composite or Hypothetical as sensitivity
Is the clinical question about tolerability? (e.g., efficacy in adherent patients)
→ Use Principal Stratum with Treatment Policy as sensitivity
Is measurement only meaningful during active treatment? (e.g., symptom scores)
→ Use While-on-Treatment (primary) with Treatment Policy as sensitivity

Crossover Handling in Open-Label Oncology Trials

Open-label oncology trials (common in late-phase or rare-disease settings) often permit crossover or secondary treatment access to control-arm patients upon progression. This creates a major OS confounding problem.

Why Crossover Matters for OS

Crossover = treatment switching after randomization, confounding the causal effect of the assigned treatment on survival
Selection bias: Patients crossing over may differ (outcome predictors) from those not crossing over
Ethical imperative: Denying effective therapy to control-arm patients at progression raises ethical concerns, but creates statistical challenge for OS

Three Methods for Crossover Adjustment

RPSFT (Rank-Preserving Structural Failure Time)

Approach: Assume counterfactual OS in control arm follows an accelerated failure time (AFT) model:

S₀*(t) = S₀(t × exp(−ψ))

where:

S₀(t × exp(−ψ)) = counterfactual survival curve for control arm if they had received treatment immediately
ψ = log(hazard ratio); estimated from the data
exp(ψ) = time acceleration factor

Mechanics:

Fit Cox model to observed data with crossover indicator
Estimate ψ by comparing rank-preserved times in treatment vs. control
Construct counterfactual curve; compare to treatment arm

Strengths:

Straightforward conceptually; single parameter ψ
Used successfully in RECORD-1 trial (2.4-month OS gain after RPSFT)
Widely accepted by FDA

Weaknesses:

AFT assumption is strong and rarely tested for validity
Sensitive to ψ estimates (small changes → large survival differences)
Can produce implausible counterfactuals (e.g., negative survival times) if model is misspecified
Assumes no unmeasured confounding before or after crossover

When to use: Informative crossover (many patients cross over late); modest crossover rates (< 50%); homogeneous treatment effect

Real-world example: RECORD-1 (everolimus, renal cell cancer) showed 2.4-month OS improvement (19% HR reduction) after RPSFT adjustment for crossover

2SRST (Two-Stage Randomization Structural Treatment Effect)

Approach: Model crossover as a "second randomization" event. Estimate:

E[OS | assigned treatment, no crossover]
E[OS | assigned control, no crossover]
E[OS | assigned control, crossed over at time t]

Weighting formula:

For each patient, weight = 1 / P(no crossover at time t | baseline covariates)

Then compare weighted arms.

Strengths:

More flexible than RPSFT; fewer distributional assumptions
Doesn't require AFT model
Can accommodate time-varying treatment effects
Easier to explain to clinicians ("control patients weighted by probability of not crossing over")

Weaknesses:

Requires careful weight stabilization (weights can become very large)
Less commonly used; less regulatory precedent
Requires sufficient sample size for accurate weight estimation
Still requires exchangeability/no unmeasured confounding

When to prefer: High crossover rates (> 50%); heterogeneous treatment effects; concern about AFT assumptions

IPCW (Inverse Probability of Censoring Weighting)

Approach: Weight each patient by the inverse probability of remaining uncrossed over, given their baseline covariates:

Weight_i(t) = 1 / P(no crossover at time t | X_i(t))

Then apply standard Cox regression or log-rank test to weighted data.

Strengths:

Directly addresses confounding by post-baseline covariates (e.g., early tumor response drives both crossover decision and OS)
Clear causal inference framework
Flexible (accommodates any functional form for crossover model)

Weaknesses:

Assumes positivity (all patients must have non-zero probability of avoiding crossover); often violated in oncology
Extreme weights when some patients are very likely to cross over
Requires careful weight stabilization and trimming

When to use: When crossover is confounded by post-baseline covariates (e.g., RECIST response, biomarkers)

Weight Stabilization and Trimming

All three methods can produce extreme weights. To address:

Stabilization:

Stabilized Weight = P(crossover | treatment assignment) / P(crossover | treatment assignment, baseline covariates)

Denominator accounts for baseline differences; numerator "marginal" weight centers on 1.

Trimming/Truncation:

Percentile trimming: Remove top and bottom 1% of weights
Fixed truncation: Cap all weights at max value (e.g., weight = 10 or 20)
Trade-off: Reduces variance but introduces bias; justify in SAP

When to Use Which Method

Scenario	Best Method	Why
Homogeneous treatment effect; modest crossover (< 30%)	RPSFT	Simplest; well-established
High crossover (> 50%); concern about AFT	2SRST	Fewer assumptions
Post-baseline confounders evident (response-driven crossover)	IPCW	Addresses post-baseline confounding
Blinded data available (early trials)	None (use hypothetical CI or censor at crossover)	Structural methods not applicable

SAP Language Templates

Template 1: Treatment Policy Strategy for Treatment Discontinuation

Estimand: Overall Survival under the treatment policy strategy.

Definition: OS is measured and reported regardless of treatment discontinuation or modification. All randomized patients will be followed for OS until death, loss to follow-up, or end of trial, per the treatment policy strategy.

Analysis Population: Intent-to-treat (ITT) population = all randomized patients.

Primary Analysis: Kaplan-Meier curves and stratified log-rank test, stratified by [randomization stratification factor].

Missing Data: Patients alive at final data cut will be censored at the date of last contact. Patients lost to follow-up will be censored at date last known alive.

Interpretation: This estimand answers the clinical question: "What is the effect of the assigned treatment on OS in the real-world setting where treatment discontinuation and modification occur?"

Template 2: Hypothetical Strategy with RPSFT Sensitivity for Subsequent Therapy

Estimand: Overall Survival in the hypothetical scenario where subsequent anticancer therapy does not confound the treatment effect.

Definition: OS is estimated under a hypothetical scenario where the treatment effect is not confounded by subsequent anticancer therapy received after progression.

Sensitivity Method — RPSFT Adjustment:

Method: Rank-Preserving Structural Failure Time (RPSFT) adjustment will be applied to account for OS in patients who cross over to the investigational therapy after receiving control therapy.

Model: The counterfactual OS in the control arm is estimated assuming an accelerated failure time (AFT) model:

S₀*(t) = S₀(t × exp(−ψ))

where ψ is the log hazard ratio, estimated from crossover patterns and survival data

Estimation: The treatment effect parameter ψ will be estimated by maximizing the rank-preserving likelihood, comparing rank-preserved failure times between treatment and control arms.

Assumptions:

The AFT model holds for the treatment effect

Randomization is preserved in the crossover subgroup

No unmeasured confounders of the relationship between crossover and survival

Secondary Sensitivity — IPCW:

If post-baseline covariates (e.g., RECIST response) are identified as confounders of the crossover decision, inverse probability censoring weighting (IPCW) will be applied as an alternative sensitivity.

Weights will be stabilized and trimmed at the 1st and 99th percentiles to ensure stability.

Reporting: RPSFT-adjusted OS curve will be compared to the unadjusted (treatment policy) OS curve. Point estimates (median OS, 1-year OS) and hazard ratios will be reported for both estimates.

Interpretation: The RPSFT estimate answers: "What would OS have been if subsequent therapy had not confounded the treatment effect?" This is presented as a sensitivity/robustness check to the primary treatment policy analysis.

Template 3: Composite Strategy for Death or Progression

Estimand: Progression-Free Survival (composite of progression or death).

Definition: PFS is defined as the time from randomization to the first occurrence of: - Radiologically confirmed disease progression per [RECIST 1.1 / mRECIST / other criteria], OR - Death from any cause, whichever occurs first.

Justification for Composite: Death before progression is a terminal event that precludes measurement of progression; therefore, death is included as a competing event in the composite endpoint.

Analysis Population: Intent-to-treat (ITT) population.

Event Definition Hierarchy: - If both progression and death occur within 14 days, the event is classified as death. - If progression occurs without prior death assessment, progression is the recorded event.

Measurement: Disease status will be assessed per protocol schedule. Deaths will be documented from hospital records, death certificates, or family contact.

Censoring: Patients without a documented PFS event will be censored at the date of the last adequate tumor assessment or end of follow-up.

Primary Analysis: Kaplan-Meier curves with log-rank test, stratified by [randomization factor].

Competing Risks: A competing risks analysis (cumulative incidence function) will be presented as a secondary analysis to distinguish death from non-cancer causes vs. progression-related death.

Interpretation: This composite endpoint reflects both efficacy (delay of progression) and tolerability (avoidance of early death), providing a clinically meaningful summary.

Regulatory Position (FDA & EMA, Post-ICH E9(R1))

FDA Guidance on Estimands in Oncology Trials

Key Principles (2019–2025):

Treatment Policy is default: FDA generally accepts treatment policy strategy as primary for oncology trials because it reflects real-world clinical use.
Hypothetical is sensitive analysis: Hypothetical estimands (e.g., RPSFT for OS) are accepted as sensitivity or supportive but rarely as primary.
Composite for terminal events: Death or other terminal events should be incorporated as composite endpoints rather than excluded.
Transparency on assumptions: Any structural assumption (AFT in RPSFT, positivity in IPCW) must be clearly stated in the SAP and sensitively explored.

Oncology-Specific Expectations:

OS is primary: OS trials should specify OS estimand (usually treatment policy).
PFS/TTP: Must specify whether discontinuation due to AE is captured (usually yes, as part of treatment policy).
Subsequent therapy: If anticipated, state upfront whether treatment policy or hypothetical (with RPSFT) is primary; FDA will request both in submission.
Crossover: In open-label trials, RPSFT or 2SRST sensitivity expected if crossover > 10–15%.

EMA Position

Similar to FDA:

Emphasizes clear estimand specification (ICH E9(R1) adoption completed 2019)
Treatment policy is pragmatic and preferred for most efficacy endpoints
Hypothetical acceptable for mechanistic questions but requires strong justification of assumptions

Differences:

More accepting of hypothetical strategy if mechanistic question is clinically relevant
Emphasizes sensitivity analyses over primary hypothetical (similar to FDA)

Controversial Choices (Risk of Rejection)

Choice	Risk	Mitigation
Principal stratum without clear clinical rationale	High (seen as excluding inconvenient data)	Prespecify stratum in protocol; justify clinical relevance
While-on-treatment as primary for OS	High (introduces informative censoring bias)	Use only for symptom/QoL endpoints; prespecify in protocol
Hypothetical without structural method (e.g., just censor at crossover)	High (loses information; biased if crossover informative)	Use RPSFT, 2SRST, or IPCW; justify choice of method
Missing structured sensitivity analyses	Medium-High	Plan at least treatment policy + hypothetical; specify which is primary
Ignoring post-baseline confounders in RPSFT	Medium (if confounders obvious in data)	Conduct IPCW sensitivity if response/biomarkers drive crossover

Clinical Interpretation Examples

Example 1: Metastatic NSCLC, Anti-PD-1 vs. Chemotherapy

Scenario:

Phase 3, open-label (blinding not feasible for immunotherapy)
OS primary endpoint
Expected 50% crossover from chemo to anti-PD-1 at progression

Estimand Choices:

Primary: Treatment policy (OS regardless of crossover)
Sensitivity: Hypothetical with RPSFT (OS adjusted for crossover effect)

SAP Specification:

"The primary analysis of OS will follow the treatment policy strategy: all randomized patients are followed for OS to death, progression, loss to follow-up, or end of trial, regardless of treatment discontinuation or crossover. Median OS and 1-year OS will be reported with 95% CIs using Kaplan-Meier estimator. A Cox proportional hazards model, stratified by [prognostic factor], will test the treatment effect. In a sensitivity analysis, an RPSFT adjustment for OS will account for crossover from chemotherapy to anti-PD-1. The treatment effect parameter ψ will be estimated via rank-preserving likelihood, and the RPSFT-adjusted OS curve will be compared to the unadjusted treatment policy curve."

Why This Works:

Treatment policy is pragmatic (reflects real-world use where crossover happens)
RPSFT sensitivity addresses the major confounding threat (crossover)
FDA/EMA accepts both for approval decisions

Example 2: Rare Pediatric Malignancy, Single-Arm to Historical Control

Scenario:

Phase 2, single-arm trial with historical control comparator
High expected early deaths (20% by 6 months) before progression can be assessed
Limited sample size (N = 50)

Estimand Choices:

Primary: Composite (death or progression as PFS)
Sensitivity: Treatment policy with censoring at loss to follow-up

SAP Specification:

"The primary efficacy endpoint is Progression-Free Survival (PFS), defined as the time from enrollment to the first occurrence of radiologically confirmed disease progression per [standard criteria] or death from any cause, whichever occurs first. Death from any cause is incorporated into the PFS composite definition because early death precludes the ability to assess progression and is a critical clinical outcome. Kaplan-Meier estimator will be used to calculate median PFS. Patients without a documented PFS event will be censored at the date of last radiological assessment or end of follow-up."

Why This Works:

Composite addresses reality (early deaths in aggressive disease)
No sensitivity analysis needed (death is clinically integral to the endpoint)
Single-arm comparison to historical control acceptable for rare disease with composite PFS

Backlinks

ICH E9(R1) Estimand Framework Intercurrent Events in Oncology Trials Sensitivity Analyses for Estimands Overall Survival (OS) Progression-Free Survival (PFS) Multiple Endpoints and Alpha Allocation Missing Data: Mechanisms, Methods, and Estimand-Driven Strategy RCT Design Fundamentals in Oncology

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, adopted November 2019; FDA adoption May 2021) Compiled from ICH E9(R1) §A.1–A.4 + estimand oncology frameworks

Last Updated: April 2026
Knowledge Base: oncology_kb
Section: Clinical Trial Design — Intercurrent Events and Estimand Strategies