Advanced Adaptive Trials

Advanced Adaptive Trial Designs in R

Overview

Advanced adaptive clinical trial designs including platform trials, basket and umbrella trials, response-adaptive randomization, multi-arm multi-stage designs, Bayesian adaptive methods, and sample size re-estimation techniques.

Platform Trials

Using adaptr Package

library(adaptr)

# Define a platform trial with multiple arms
setup 
  setup_trial_binom(
    highest_is_best = TRUE,
    soften_power = 0.5                    # Softening for allocation
  )

# Add arm dropping rules
setup 
  add_arm(
    arm = "Arm_C",
    true_y = 0.50,
    start_look = 3              # Add at third interim
  )

# Simulate
sims_platform  p0
fit_mem$post_prob

# Cluster map (which baskets share information)
plot(fit_mem, type = "cluster")

Hierarchical Model for Baskets

library(basket)

# Full Bayesian hierarchical model
fit_hier  p0) > 0.975
  qc = 0.10,              # Futility: P(rate > p0) < 0.10
  lower.tail = FALSE
)

# Operating characteristics
oc <- oc2S(
  prior_treatment = robust_prior,
  prior_control = robust_prior,
  n1_treatment = 30,      # Stage 1 treatment
  n1_control = 30,        # Stage 1 control
  n2_treatment = 30,      # Stage 2 treatment
  n2_control = 30,        # Stage 2 control
  decision = decision
)

# Plot OC curves
plot(oc)

# Type I error and power
summary(oc)

Group Sequential Designs with rpact

O'Brien-Fleming Design

library(rpact)

# O'Brien-Fleming group sequential design
design_of <- getDesignGroupSequential(
  kMax = 3,                   # Number of stages
  alpha = 0.025,              # One-sided alpha
  beta = 0.20,                # Type II error
  sided = 1,
  typeOfDesign = "OF",
  informationRates = c(0.33, 0.67, 1.0)
)

summary(design_of)
plot(design_of)

# Boundaries
design_of$criticalValues    # Z-score boundaries
design_of$alphaSpent        # Cumulative alpha spent

Sample Size Calculation

library(rpact)

# Sample size for survival endpoint
sample_size <- getSampleSizeSurvival(
  design = design_of,
  lambda1 = log(2) / 24,      # Control median = 24 months
  lambda2 = log(2) / 36,      # Treatment median = 36 months (HR = 0.67)
  accrualTime = 24,           # Accrual period
  followUpTime = 12,          # Additional follow-up
  dropoutRate1 = 0.05,        # Control dropout
  dropoutRate2 = 0.05,        # Treatment dropout
  allocationRatioPlanned = 1
)

summary(sample_size)

# Events and sample size at each look
sample_size$eventsPerStage
sample_size$numberOfSubjects

Interim Analysis

library(rpact)

# Perform interim analysis
interim <- getAnalysisResults(
  design_of,
  dataInput = getDataset(
    n = c(100, 100),          # Cumulative n by stage
    events = c(40, 80),       # Cumulative events by stage
    logRanks = c(2.1, 2.8)    # Log-rank Z statistics
  )
)

summary(interim)

# Can the trial stop?
interim$finalStage           # Final stage reached?
interim$futilityStop         # Stopped for futility?
interim$rejectAtFinalStage   # Rejected at final analysis?

Sample Size Re-Estimation

Blinded SSR

library(rpact)

# Sample size re-estimation based on interim data
ssr <- getSampleSizeReestimation(
  design_of,
  stageResults = interim,
  conditionalPower = 0.80     # Target conditional power
)

summary(ssr)
ssr$sampleSizeNew             # New sample size recommendation

Unblinded SSR

library(rpact)

# Conditional power at interim
cp <- getConditionalPower(
  design = design_of,
  stage = 2,
  stageResults = interim,
  nPlanned = c(50, 50),       # Planned future n per arm
  assumedEffect = 0.67        # Assumed treatment effect (HR)
)

summary(cp)

# If CP too low, calculate required sample size
if (cp$conditionalPower < 0.50) {
  # Increase sample size
  new_n <- getSampleSizeReestimation(design_of, interim, conditionalPower = 0.80)
}

Graphical Multiplicity with gMCP

library(gMCPLite)

# Define hypothesis graph
# H1, H2: Primary endpoints; H3, H4: Secondary endpoints
m <- matrix(
  c(0, 0.5, 0.5, 0,
    0.5, 0, 0, 0.5,
    0.5, 0, 0, 0.5,
    0, 0.5, 0.5, 0),
  nrow = 4, byrow = TRUE
)

weights <- c(0.5, 0.5, 0, 0)  # Initial alpha allocation

# Create graph
graph <- gMCP::matrix2graph(m, weights)
nodeNames(graph) <- c("H1_Primary", "H2_Primary", "H3_Secondary", "H4_Secondary")

# Plot
plot(graph)

# Test with p-values
pvalues <- c(0.01, 0.03, 0.02, 0.04)
result <- gMCP::gMCP(graph, pvalues, alpha = 0.025)
print(result)

# Which hypotheses rejected?
result@rejected

Key Packages Summary

| Package | Purpose | |---------|---------| | adaptr | Platform trial simulation | | basket | Basket trial analysis | | MAMS | Multi-arm multi-stage designs | | rpact | Group sequential and adaptive designs | | RBesT | Bayesian evidence synthesis | | gMCPLite | Graphical multiplicity procedures | | gsDesign | Group sequential designs | | gsDesign2 | Enhanced group sequential | | Mediana | Clinical trial simulations |

Best Practices

1. Pre-specification: Define adaptation rules before trial starts
2. Type I error: Ensure strong control under all adaptations
3. Operating characteristics: Simulate extensively under various scenarios
4. Blinding: Maintain blinding where possible during adaptations
5. Documentation: Document all decision rules in protocol
6. Regulatory: Engage regulators early for complex adaptive designs
7. Implementation: Plan for operational complexity of adaptations
8. Analysis: Plan for potential biases from adaptations

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: choxos
• Source: choxos/BiostatAgent
• License: MIT

Install and usage instructions live in the source repository linked above.