Adaptive Trial Designs in Oncology

Definition

An adaptive design is a clinical trial that allows prospectively planned modifications to the design, sample size, population, randomization ratio, or statistical analysis based on accumulating interim data, while preserving the type I error rate and maintaining statistical power. Adaptations are determined by pre-specified rules codified in the Statistical Analysis Plan (SAP) before unblinding.

Regulatory perspective (FDA 2019, Final): Adaptive designs are permitted for all phases (exploratory, confirmatory); the key requirement is pre-specification and rigorous type I error control.

Regulatory perspective (ICH E20 2025, DRAFT Step 2b): Formalizes adaptive design principles globally; stricter requirements for simulation documentation and sensitivity analyses; introduces Bayesian adaptive frameworks as co-equal to frequentist conditional error function approaches.

FDA 2019 Classification: Well-Understood vs Less Well-Understood Adaptive Designs

Well-Understood Adaptive Designs (Lower Regulatory Burden)

Characteristics: Extensive experience; clear type I error control; limited complexity

Examples:

Group Sequential Designs (GSD)
- Interim efficacy/futility stopping with pre-specified boundaries (O'Brien-Fleming, Pocock)
- No sample size re-estimation; no population changes
- Type I error controlled by spending functions
- Regulatory burden: Low; standard analysis
- R packages: rpact, gsDesign2
Blinded Sample Size Re-Estimation
- Recalculate sample size at interim based on nuisance parameters only (event rate, dropout rate)
- Do NOT unblind treatment effect
- Preserves type I error under standard conditional error function
- Example: Event rate lower than expected → increase N proportionally
- Regulatory burden: Low; straightforward inflation factor or conditional error analysis
- R packages: rpact, adaptIVPT
Planned Interim Analysis with Futility Only
- Test for futility (conditional power <20%) at interim
- Efficacy tested only at final analysis
- Type I error = α (no inflation)
- Regulatory burden: Minimal

Less Well-Understood Adaptive Designs (Higher Regulatory Burden)

Characteristics: Complex type I error control; limited historical experience; requires simulation validation

Examples:

Unblinded Sample Size Re-Estimation
- Re-estimate N based on observed treatment effect at interim
- Adaptive combination test (p₁, p₂, correlation structure) required to preserve α
- Risk: If treatment effect weak, adaptive increase in N + unblinded data → type I error inflation
- Control: Conditional error function approach (Cui, Hung, Wang 1999)
- Regulatory burden: HIGH; requires simulation proof of type I error control
- R packages: rpact (conditional error), MAMS
Population Enrichment (Adaptive Subgroup Selection)
- Full population enrolled initially
- Interim analysis: If pre-specified biomarker/subgroup shows superior effect, adapt to enrich that population in stage 2
- Risk: Multiple testing across biomarkers; inflation if not pre-specified
- Control: Hierarchical gatekeeping or Bayesian methods; pre-specification of biomarkers and decision rules
- Example: PD-L1+ vs PD-L1- cohorts in immunotherapy trial; adapt if PD-L1+ shows benefit
- Regulatory burden: VERY HIGH; FDA requires detailed biomarker rationale, pre-specification, simulation
Adaptive Randomization (Response-Adaptive, Covariate-Adaptive)
- Randomization probabilities change during trial based on:
- Response-adaptive: Interim efficacy/safety data (e.g., favor arm with higher response rate)
- Covariate-adaptive: Balance baseline covariates in real-time
- Risk: Bias in estimation; loss of blinding if observed (e.g., more patients to one arm = possible unblinding)
- Control: Analysis must adjust for adaptive allocation; conditional error function + sensitivity analysis
- Regulatory burden: VERY HIGH; FDA skeptical; requires rigorous justification and simulation
- R packages: adaptIVPT, specialized Bayesian software
Seamless Phase 2/3 Designs
- Single trial spanning dose-ranging (Phase 2) and efficacy (Phase 3)
- Interim: Select optimal dose(s) based on Phase 2 data
- Stage 2: Confirm efficacy at selected dose(s) in larger population
- Risk: Dose selection based on same population; if selection wrong, phase 3 underpowered; multiplicity across doses
- Control: Pre-specification of dose selection criteria; hierarchical testing or closed testing procedures; often Bayesian
- Example: KEYNOTE-407 (pembrolizumab in squamous NSCLC) used adaptive design for dose confirmation
- Regulatory burden: HIGH; requires simulation and pre-specification

ICH E20 (2025 Draft, Step 2b): Key Differences from FDA 2019

What ICH E20 Adds to FDA 2019

Formal Bayesian Adaptive Framework
- FDA 2019: Mentions Bayesian methods; primarily focuses on frequentist conditional error function
- ICH E20 DRAFT: Elevates Bayesian adaptive designs to co-equal status with frequentist approaches
- Allows Bayesian response-adaptive randomization with proper type I error control via posterior predictive probability
- Recognizes informative priors from historical data as valid for confirmatory trials (if pre-specified)
Stricter Simulation Requirements
- FDA 2019: Simulation recommended for complex designs; not always required
- ICH E20 DRAFT: Mandatory simulation for all adaptive designs with:
- ≥10,000 replicates under null (type I error)
- ≥1,000 replicates under alternative (power)
- Scenario coverage: All plausible parameter combinations (not single-point estimates)
- Sensitivity analysis: MNAR (missing not at random) robustness if applicable
Explicit Guidance on IDMC (Independent Data Monitoring Committee) Role
- FDA 2019: IDMC role implied; details sparse
- ICH E20 DRAFT: Detailed specifications:
- IDMC must have statistical expertise in adaptive design methodology
- Interim futility decisions must be based on pre-specified conditional power thresholds (not subjective judgment)
- Adaptive allocations (randomization ratio change, population switch) require IDMC recommendation + sponsor transparency
Intercurrent Events and Missing Data Under Adaptation
- FDA 2019: Links to ICH E9(R1) but doesn't address how adaptations interact with intercurrent events
- ICH E20 DRAFT: Explicit guidance:
- If population enrichment occurs, estimand must be re-defined (principal stratum, hypothetical scenario, etc.)
- If randomization adapts, missing data strategy may need re-specification in the enriched population
- Example: If trial enriches to biomarker+ patients at interim, primary estimand shifts to biomarker+ population; missing data assumption (MAR vs MNAR) must be re-justified for new population
Historical Data and Borrowing
- FDA 2019: Cautious on external data; generally recommends prospective trials
- ICH E20 DRAFT: Formalizes use of historical data borrowing via:
- Informative priors (Bayesian)
- Power priors (weighted historical information)
- Requires prior-data conflict assessment (Bayesian predictive checks)
- Discordance plan: What if historical data conflict with interim findings?
Timing and Pre-specification in SAP
- FDA 2019: Pre-specification recommended; some flexibility for modifications via Type C meeting
- ICH E20 DRAFT: Stricter: All adaptive rules, decision thresholds, and conditional power/futility criteria must be in SAP before any unblinding
- Modifications post-IND or after interim only allowed under exceptional circumstances (with written FDA justification)

Sample Size Re-Estimation: Blinded vs Unblinded

Blinded Sample Size Re-Estimation (Well-Understood)

When permissible: Adjust N based on nuisance parameters (event rate, dropout, variance) observed at interim, without unblinding treatment effect.

Procedure:

At interim (e.g., 50% information), assess event rate or variance in pooled population (all arms combined)
Compare to original assumption
Recalculate sample size: N_new = N_original × (observed / assumed)²
Increase enrollment if needed; continues blinded
Final analysis: Combined test across both stages

Type I Error Control: Preserved automatically under blinded SSR (conditional error function not required)

Example (Oncology):

# Original assumption: median OS = 12 months (λ = 0.0578/month)
# At interim: observed pooled event rate = 0.035/month (higher than assumed)
# This means fewer events will accrue → longer trial duration needed
# Recalculate: N_new = N_original × (0.0578 / 0.035)² ≈ 1.37 × N_original

R Package: rpact supports blinded SSR with inflation factor calculation

Unblinded Sample Size Re-Estimation (Less Well-Understood)

When permissible: Recalculate N based on observed treatment effect (e.g., observed HR) at interim.

Risk: Observing weak effect → unblinded increase in N → temptation to declare success if final p-value barely crosses α, even though weak effect was true all along.

Type I Error Control Required: Conditional Error Function (Cui, Hung, Wang 1999)

Procedure:

At interim, unblind and observe:
- Z₁ = interim test statistic (standardized log-rank, for example)
- p₁ = interim p-value
Calculate conditional error (or equivalently, conditional power):
- CP = P(reject at final | interim data)
- If CP > 80%, trial likely to succeed; maybe stop early for efficacy
- If CP < 20%, trial futile; stop for futility
- If 20% < CP < 80%, continue with re-estimated N
Pre-specify decision boundaries (α-spending function) that account for the interim look
Final analysis combines Z₁ and Z₂ via combination test:
- Z_combined = w₁ × Z₁ + w₂ × Z₂ (weights pre-specified)
- Or: Inverse normal method: p_final = Φ(w₁ × Φ⁻¹(p₁) + w₂ × Φ⁻¹(p₂))

Type I Error Control: Guaranteed by proper weighting and spending function choice (requires simulation validation)

Example SAP Language:

At interim analysis (50% information), if observed treatment effect is weak 
(HR > 0.80), the trial will be re-powered to detect HR = 0.80 instead of 
the original 0.70. The conditional error function approach (Cui, Hung, Wang) 
will be applied. Final analysis will use inverse normal combination method 
with weights: w₁ = 0.5, w₂ = 0.5 (symmetric).

Type I error control verified via simulation (10,000 replicates under null 
hypothesis) showing α-level ≤ 0.025.

R Packages:

rpact: Conditional error function, design, and analysis functions
MAMS: Multi-arm, multi-stage adaptive designs with SSR

Adaptive Randomization

Response-Adaptive Randomization (RAR)

Mechanism: Randomization probabilities shift during trial based on interim efficacy or safety.

Example:

Initially: 1:1 randomization (50% to each arm)
At interim: Observe arm A response rate = 60%, arm B = 40%
Adapt: Shift to 65:35 (favor responding arm A)
Rationale: Ethical (more patients get better treatment); efficiency (fewer failures)

Risk:

Bias in estimation: If allocation favors arm A, final effect estimate may be inflated
Type I error inflation: If decision rule based on interim p-value, unblinded re-randomization can amplify significance
Unblinding risk: Unequal randomization visible to site staff; can compromise blinding

Type I Error Control: Requires conditional error function + sensitivity analysis

Regulatory stance (FDA 2019): Possible, but skeptical; requires:

Robust justification (ethical benefit clearly outweighs complexity)
Pre-specified adaptation rule (not data-driven)
Simulation showing type I error control
Sensitivity analysis (assume rule failed; what if allocation not adaptive?)

Regulatory stance (ICH E20 DRAFT): Opens door to Bayesian response-adaptive randomization using posterior probabilities:

Randomization probability ∝ posterior P(arm is best | interim data)
Type I error controlled via Bayesian predictive probability framework
More transparent than frequentist RAR

R Packages:

adaptIVPT: Inverse probability weighting for RAR
Bayesian software: rstan, Stan models for Bayesian RAR

Covariate-Adaptive Randomization

Mechanism: Real-time balancing of baseline covariates (age, disease stage, biomarker status) across arms.

Example:

PD-L1 status and ECOG performance status are critical confounders
Traditional 1:1 randomization may imbalance these in small trials
Covariate-adaptive (e.g., minimization, stratified permuted blocks): Ensure balance in real-time

Advantages:

Reduces confounding; increases precision
No impact on type I error (statistical nuisance)
Blinding preserved (allocation method invisible to patients)

Regulatory stance: Well-understood; FDA 2019 accepts covariate-adaptive randomization routinely.

Minimal Requirements:

Pre-specify covariates and balancing algorithm
Document in SAP
No simulation needed (deterministic allocation)

R Packages:

blockrand: Stratified permuted blocks
minimise: Minimization algorithms

Seamless Phase 2/3 Designs and Population Enrichment

Seamless Phase 2/3 Design

Structure:

Stage 1 (Phase 2): Dose-ranging study in smaller population (e.g., n=100)
Objective: Select optimal dose based on efficacy/tolerability
Primary endpoint: Response rate or dose-limiting toxicity
Stage 2 (Phase 3): Efficacy confirmation at selected dose in larger population (e.g., n=300 additional)
Objective: Confirm efficacy of chosen dose vs. control
Primary endpoint: OS or PFS

Advantages:

Time and cost savings (single trial instead of two)
Expedited development pathway

Risks:

Dose selection bias: If Phase 2 population is small or unrepresentative, selected dose may not be optimal for Phase 3 population
Multiplicity: If testing multiple doses in Phase 2, selection introduces multiple comparisons
Type I error inflation: If final analysis uses same Phase 2 data to both select dose and test efficacy

Type I Error Control:

Closed testing procedures (graphical methods, gatekeeping)
Hierarchical hypothesis structure:
- Primary: Test efficacy of selected dose (Phase 3 data alone)
- Secondary: Test dose A, then dose B, etc. (Phase 2 data, closed testing)
Bayesian approach: Informative prior on selected dose based on Phase 2 data; Phase 3 likelihood-based update

Regulatory stance (FDA 2019): Acceptable; requires clear pre-specification of dose selection rule.

Regulatory stance (ICH E20 DRAFT): Favors structured Bayesian approach for dose selection and confirmation.

Example (Real Trial): KEYNOTE-407 (pembrolizumab + chemotherapy in squamous NSCLC)

Phase 2: Pembrolizumab 10 mg/kg Q2W selected
Phase 3: Confirmed 10 mg/kg as optimal (vs. chemotherapy alone)
Design: Seamless, with dose selection at interim

Population Enrichment Adaptation

Mechanism: Trial starts with broad population; at interim, focus on biomarker-defined subgroup (e.g., PD-L1 high, BRCA mutant).

Rationale:

Biomarker effect uncertain at trial start (continuous response across subgroups)
Interim data reveals biomarker signal
Efficiency: Enrich to biomarker+ patients if they show benefit

Procedure:

Stage 1: Enroll all-comers (n=200) with baseline biomarker assessment
Interim Analysis (after ~100 events):
- Compare treatment effect in biomarker+ vs. biomarker- subgroups
- Decision rule (pre-specified):
- If biomarker+ subgroup shows superior response (e.g., HR < 0.65), enrich to biomarker+ only
- Otherwise, continue with all-comers
Stage 2: Enroll additional biomarker+ patients (n=200 more), if enriched

Type I Error Control:

Challenge: Testing in one population (all-comers), then switching to subgroup → multiplicity
Solution: Hierarchical testing (test biomarker+ efficacy only if overall effect significant) OR closed testing + Holm-Bonferroni correction
Bayesian alternative: Mixture prior reflecting uncertainty; posterior probability of biomarker+ benefit guides decision

Regulatory stance (FDA 2019): Requires:

Clear clinical/biological rationale for biomarker
Pre-specified enrichment rule (decision threshold, e.g., "enrich if HR < 0.65 in biomarker+ subgroup")
Hierarchical testing or closed testing to control type I error
Simulation validation

Regulatory stance (ICH E20 DRAFT): Formalizes via adaptive subgroup selection framework:

Prior on biomarker effect specified (informative or flat)
Interim decision rule based on posterior probability of benefit in each subgroup
Phase 3 efficacy estimated in enriched population; adjusted for prior data borrowing

Estimand Impact:

Original estimand (all-comers, treatment policy): "Effect of assignment to treatment in overall population"
Enriched estimand (biomarker+, hypothetical): "Effect in biomarker+ subgroup, hypothetically continuing treatment"
SAP must specify: How does enrichment change the estimand? (ICH E9(R1) implications)

Type I Error Control: Conditional Error Function and Combination Tests

Conditional Error Function (CEF) Approach

Principle: At interim, compute the conditional error probability (probability of rejecting H₀ at final, given current data).

Mathematical framework (Cui, Hung, Wang 1999):

Stage 1: Collect n₁ observations; compute test statistic Z₁, p-value p₁
Stage 2: Collect n₂ observations; compute Z₂, p₂

Conditional Error: α_c = P(reject H₀ | interim data)
                       = P(Z₂ > z(α_c | Z₁))

If interim decision is "continue to stage 2", the spending boundary at stage 2 is 
adjusted so that if Z₂ crosses it, the combined test will reject H₀ with overall 
type I error = α.

Advantages:

Flexible: Allows sample size change, population switch, test change
Efficient: Weak interim findings can be re-powered
Transparent: Decision rules tied to conditional power

Implementation (Inverse Normal Method):

Z_combined = w₁ × Z₁ + w₂ × Z₂  (weights w₁, w₂ pre-specified)

Critical value at stage 2: c₂ = (z_α - w₁ × Z₁) / w₂

If Z₂ > c₂, reject H₀ (at stage 2)

Example (Oncology):

Stage 1: 200 events observed; Z₁ = 1.5 (p₁ = 0.067, non-significant)
Interim decision: Weak signal; re-estimate sample size for final stage 2
Pre-specified weights: w₁ = 0.4, w₂ = 0.6 (asymmetric; more weight on stage 2)
Stage 2: Collect 200 more events; Z₂ = 2.1
Z_combined = 0.4 × 1.5 + 0.6 × 2.1 = 1.86
If critical value ≤ 1.86, reject H₀ at overall α = 0.025

R Packages:

rpact: Conditional error functions, design, analysis, inflation factors
MAMS: Multi-arm, multi-stage with conditional error control

Combination Tests (Fisher, Inverse Normal)

Fisher's Method (for p-values):

Test statistic: T_Fisher = -2 × [ln(p₁) + ln(p₂)]
Null distribution: χ² with 4 degrees of freedom

Advantage: Intuitive; based on p-values Disadvantage: Less flexible for unequal information weights

Inverse Normal Method (for z-statistics):

Z_combined = (w₁ × Z₁ + w₂ × Z₂) / √(w₁² + w₂²)
Null distribution: N(0, 1) [standard normal]

Advantage: Flexible; allows asymmetric weights for unequal stages Disadvantage: Assumes normally distributed test statistics

Pre-specification in SAP:

"The primary analysis will use the inverse normal combination test with 
weights w₁ = 0.5, w₂ = 0.5 (symmetric, equal information at each stage).

Type I error control verified via simulation:

- 10,000 replicates under H₀ (HR = 1.0)
- Estimated α = 0.0248 (95% CI: 0.0210–0.0290)
- Conclusion: Type I error ≤ 0.025 ✓
"

Simulation Requirements per ICH E20 (DRAFT)

Mandatory Elements

1. Simulation scope (DRAFT requirement, stricter than FDA 2019):

≥10,000 replicates for type I error (null hypothesis true)
≥1,000 replicates for power (alternative hypothesis true)
All plausible scenarios: Base-case, optimistic, pessimistic parameter values
No single-point estimates allowed; must explore parameter ranges

2. Data-generating model specification:

Piecewise hazard (if non-proportional hazards)
Event rate, dropout rate, enrollment rate
Interim timing (% information)
Any adaptive rule parameters (e.g., conditional power threshold for futility)

3. Adaptive rule validation:

Simulate the exact adaptive rule (blinded SSR, unblinded SSR, enrichment, RAR)
Compute conditional power at interim under different scenarios
Verify futility/efficacy stopping boundaries preserve α

4. Type I error control proof:

Show estimated α ≤ 0.025 (with 95% CI) under null
If multiple adaptive looks or interim decisions, confirm per-family error rate (FWER) controlled

5. Sensitivity analyses (DRAFT):

MNAR (Missing Not At Random): If dropout mechanism may be informative, simulate under MNAR assumptions (delta adjustment, reference-based imputation)
Alternative adaptive rules: If primary rule uncertain, test robustness to rule variation
Dose selection bias: For seamless Phase 2/3, simulate impact of dose selection on final power

6. Operating characteristic reporting:

Expected sample size, # events, trial duration under each scenario
Probability of early stopping (efficacy, futility) at interim
Conditional power distribution at interim
Power by subgroup (if enrichment adaptive)

Pre-Specification and FDA Communication

SAP Requirements (FDA 2019 + ICH E20 DRAFT)

Must be specified BEFORE any unblinding:

Adaptive rule(s): Exact decision criterion "At interim analysis (after 100 events), if conditional power < 20%, trial will be stopped for futility. Conditional power is calculated as P(Z_final > 1.96 | interim data, assumed HR = 0.70 under alternative)."
Type I error control method: "Inverse normal combination test with weights w₁ = 0.5, w₂ = 0.5. Type I error verified via simulation showing α ≤ 0.025."
Simulation assumptions: Document data-generating model completely "Event rate: λ_control = 0.05/month; HR = 1.0 (null), 0.70 (alternative); Dropout: 1%/month; Enrollment: 20/month; Interim timing: 50% events"
Decision boundaries: Efficacy and futility boundaries at interim "Efficacy: None (continue to final). Futility: Stop if conditional power < 20%. Final: Reject H₀ if Z_combined > 1.96 (α = 0.025, one-sided)"

FDA Communication (Type C Meeting)

When to request feedback:

Complex adaptive design (less well-understood category)
Novel adaptive rule not extensively published
Uncertainty about regulatory acceptability

Deliverables for meeting:

Study protocol (draft)
SAP (detailed adaptive rules + simulation)
Simulation report: 10,000+ replicate results showing type I error control
Sensitivity analyses (alternative rules, MNAR robustness)
Literature references supporting approach

FDA Response: Type C meeting concludes with:

"Acceptable" (design approved; proceed as planned)
"Continue discussion" (minor modifications needed)
"Not acceptable" (major redesign required)

R Packages for Adaptive Design Implementation

rpact: Comprehensive Adaptive Design Framework

Installation: install.packages("rpact")

Key functions:

getDesignGroupSequential(): Group sequential design (O'Brien-Fleming, Pocock)
getDesignAdaptive(): Adaptive sample size re-estimation with conditional error
getDesignConditionalDARTPower(): Conditional power assessment
getAnalysisResults(): Analyze trial data under adaptive design
getInverseNormalCombinationTest(): Inverse normal combination test

Example (Unblinded SSR):

library(rpact)

# Design: Unblinded SSR with conditional error function
design <- getDesignAdaptive(
  alpha = 0.025,
  beta = 0.2,
  informationRates = c(0.5, 1.0),  # Interim at 50% info
  typeOfDesign = "conditionalErrorFunction",
  typeBetaSpending = "bsf",         # Beta-spending function
  gammaA = 1.0                        # Power parameter
)

# Simulate under adaptive rule
sim <- getSimulationMultiArmSurvival(
  design = design,
  n = 400,
  lambda = 0.05,              # Control event rate
  theta = log(0.65),           # Treatment effect (HR)
  dropoutRate = 0.01,
  maxNumberOfIterations = 10000
)

summary(sim)

adaptIVPT: Adaptive Designs with Inverse Probability Weighting

Installation: GitHub (Merck package)

Purpose: Response-adaptive randomization with bias correction

Key function: adaptIVPT() implements inverse probability weighting (IPW) to adjust for adaptive allocation bias

Use case: If response-adaptive randomization deployed, IPW ensures unbiased treatment effect estimate

MAMS: Multi-Arm, Multi-Stage Adaptive Designs

Installation: install.packages("MAMS")

Key features:

Designs for multiple treatment arms with interim arm dropping
Sample size re-estimation
Conditional error function with combination tests
Flexible efficacy/futility boundaries

Example (Multi-arm design):

library(MAMS)

# Design: 3 arms (A, B, control) with interim arm dropping
design <- mams(
  nMat = matrix(c(100, 100, 100, 200, 200, 200), nrow = 2, byrow = TRUE),
  alpha = 0.025,
  beta = 0.2,
  r = 2,                  # 2 stages
  r0 = 2,                 # 2 treatment arms initially
  rar = TRUE              # Response-adaptive randomization
)

# Analyze using combination test

Backlinks

Source: FDA Guidance "Adaptive Designs for Clinical Trials of Drugs and Biologics" (Final, November 2019); ICH E20 "Adaptive Design Clinical Trials" (DRAFT Step 2b, June 2025 revision pending); literature on conditional error functions, combination tests, Bayesian adaptive designs
Status: FDA 2019 guidance — Final; ICH E20 — DRAFT (Step 2b, June 2025 revision in progress; flagged as DRAFT per user instruction)
Compiled from: FDA 2019 adaptive guidance, ICH E20 draft documents, simtrial/rpact documentation, published literature on conditional error functions (Cui, Hung, Wang 1999; Jennison & Turnbull 2000)