Bayesian Adaptive Randomization: A Clinical Trial Worked Example

Introduction

Traditional randomized clinical trials generally specify treatment allocation probabilities before the trial begins. For example, in a two-arm study, patients may be randomized 1:1:

This approach is simple, transparent, and statistically well understood.

But suppose accumulating data suggest that Treatment A is substantially more effective than Treatment B. A conventional fixed randomization scheme continues assigning patients to both treatments with the same probabilities.

Bayesian adaptive randomization takes a different approach. The probability of assigning the next patient to each treatment can change as evidence accumulates.

The design therefore combines two components:

• Bayesian inference: update what is believed about treatment effects.
• Adaptive randomization: use those updated beliefs to determine future allocation probabilities.

Fixed Randomization vs. Adaptive Randomization

Consider a two-arm clinical trial comparing:

• Control: \(C\)
• Experimental treatment: \(E\)

Under conventional 1:1 randomization:

These probabilities remain constant throughout the trial.

Under adaptive randomization, the probabilities might evolve:

The actual allocation probabilities depend on the prespecified adaptive algorithm. They are not simply chosen by investigators after observing the data.

Why Use a Bayesian Framework?

Bayesian inference provides a natural mechanism for continuously updating beliefs about unknown treatment parameters.

Let:

and:

Before observing the trial data, we specify prior distributions:

After observing responses, Bayes' theorem updates these distributions to posterior distributions.

The posterior distribution therefore incorporates both:

• Prior information
• Current trial data

Bayesian Updating for a Binary Endpoint

For a binary response endpoint, the Beta-Binomial model is especially convenient. Suppose Treatment \(E\) has:

• \(x_E\) responses
• \(n_E-x_E\) nonresponses

with prior:

The likelihood is:

The posterior distribution is:

This is the conjugate Beta-Binomial model.

The Posterior Probability of Treatment Superiority

The quantity that often drives Bayesian adaptive randomization is the posterior probability that one treatment is better than another.

For example:

If:

the posterior evidence provides little reason to favor either treatment.

If:

there is substantial posterior probability that the experimental treatment has a higher response probability.

If:

the posterior evidence is very strong in favor of the experimental treatment.

Posterior Probability Is Not the Same as a P-Value

This distinction is fundamental.

A Bayesian probability such as:

That is conceptually different from a frequentist p-value.

From Posterior Probability to Allocation Probability

The posterior probability \(q\) does not automatically determine the randomization probability.

A separate allocation function must be specified.

One simple approach is:

Under this rule, if:

then:

This is sometimes called a probability-of-superiority allocation rule.

Why Direct Posterior Allocation Can Be Aggressive

Suppose the first few patients happen to respond well to Treatment E. The posterior probability \(q\) may become large even though the sample size is still small.

If the allocation probability is set directly equal to \(q\), the design may quickly assign most future patients to E.

That can create a feedback loop:

For this reason, practical Bayesian adaptive designs often use safeguards rather than allowing allocation probabilities to become arbitrarily extreme.

Common Allocation Rules

There is no single universally correct Bayesian adaptive-randomization formula. Common approaches include:

A Practical Safeguard: Allocation Bounds

A simple safeguard is to impose:

For example, with:

the experimental treatment can receive no less than 20% and no more than 80% of allocations.

The rule becomes:

and:

This prevents the algorithm from effectively eliminating one arm from the trial based on limited information.

A Second Safeguard: Burn-In Randomization

Another common strategy is to begin with fixed randomization.

For example, the first 40 patients might be randomized 1:1.

Adaptive randomization begins only after sufficient information has accumulated.

A Complete Worked Example

Consider a randomized Phase II trial comparing an experimental treatment with control. The primary endpoint is binary response.

Suppose the investigators begin with independent uniform priors:

The prior mean for each treatment is:

The prior therefore expresses symmetry between the two treatment arms.

Initial Randomization

Suppose the first 20 patients are randomized equally:

After the first 20 patients, suppose the observed responses are:

The observed response rates are therefore:

The experimental treatment currently appears more promising, but Bayesian inference accounts for the uncertainty caused by the small sample size.

Step 1: Update the Experimental Posterior

The experimental prior is:

There are seven responses and three nonresponses. Therefore:

so:

The posterior mean is:

The posterior mean is approximately 66.7%.

Step 2: Update the Control Posterior

For control:

There are four responses and six nonresponses. Therefore:

so:

The posterior mean is:

The posterior mean is approximately 41.7%.

Step 3: Compare the Posterior Distributions

We now have:

The Bayesian quantity of interest is:

This is the posterior probability that the experimental treatment has a higher response probability than control.

For these posterior distributions, numerical integration or Monte Carlo simulation can be used to calculate \(q\).

A convenient Monte Carlo calculation is:

set.seed(123)

M <- 1000000

p_E <- rbeta(M, 8, 4)
p_C <- rbeta(M, 5, 7)

q <- mean(p_E > p_C)

q

For this example, the posterior probability will be approximately:

The exact Monte Carlo value will vary slightly with the random seed and number of simulations.

Step 4: Convert Posterior Evidence into Allocation Probability

Suppose the adaptive rule is:

If:

then the next patient would be randomized with approximately:

Thus, the accumulating evidence has changed the randomization probabilities from 50:50 to approximately 90:10.

This is the basic mechanism of Bayesian response-adaptive randomization.

Applying an Allocation Bound

Now suppose the protocol specifies that neither arm can receive less than 20% of allocations.

The unbounded probability is:

The bounded rule is:

Therefore:

The evidence favors the experimental treatment, but the algorithm retains some probability of assigning patients to control.

Why Continue Randomizing to Control?

At first glance, assigning 20% of patients to an apparently inferior treatment may seem inefficient.

But the control arm remains important for several reasons.

• The posterior comparison depends on information from both arms.
• Early treatment differences may be exaggerated by random variation.
• Continued control observations improve estimation of the comparator.
• Clinical trials generally require a defensible comparison against a control.
• Extreme allocation probabilities can reduce inferential efficiency.

Adaptive randomization therefore involves a tradeoff between allocating more patients toward treatments that appear promising and preserving information about the comparison.

Bayesian Updating Is Sequential

Suppose the next patient is assigned to Treatment E and responds. The posterior changes immediately:

If instead that patient does not respond:

Every observed outcome therefore changes the posterior distribution.

The next allocation probability can consequently change as well.

Posterior Probability vs. Posterior Mean

Several different Bayesian quantities can be used to drive adaptive allocation. For example, investigators might use:

or the posterior mean difference:

These are not equivalent.

The posterior probability answers:

"What is the probability that E is better than C?"

The posterior mean difference answers:

"What is the expected difference in response probabilities?"

The appropriate quantity depends on the adaptive utility and decision rule specified in the trial design.

Expected Treatment Effect

Suppose the treatment effect is defined as a risk difference:

The posterior distribution of \(\Delta\) can be obtained by sampling from the two posterior distributions:

and calculating:

The resulting Monte Carlo sample approximates the posterior distribution of the treatment effect.

Posterior Probability of a Clinically Meaningful Benefit

A particularly useful Bayesian decision quantity is the probability that the treatment benefit exceeds a clinically meaningful threshold.

Suppose the clinically meaningful improvement is:

Then we might calculate:

This is different from merely asking whether:

A treatment that improves response by 1 percentage point may be statistically better in a Bayesian posterior sense but clinically unimportant.

Three Possible Bayesian Allocation Rules

Softmax Allocation

Another flexible allocation mechanism is the softmax function. Suppose each treatment has a posterior expected utility:

The allocation probability for E can be defined as:

where \(\lambda\) controls how strongly the design responds to differences in posterior utility.

If \(\lambda\) is small, allocation probabilities remain relatively balanced.

If \(\lambda\) is large, the algorithm allocates much more strongly toward the treatment with greater posterior utility.

Interpreting the Temperature Parameter

The parameter \(\lambda\) can be viewed as controlling the aggressiveness of adaptation.

This provides a convenient way to tune the balance between learning and preferential allocation.

Randomization Is Still Random

An adaptive allocation probability of 80% does not mean that the next patient must receive the experimental treatment.

It means:

The actual assignment is still random.

For example, over 10 future patients, an 80% allocation probability does not guarantee exactly eight experimental assignments.

The realized number can be 6, 7, 8, 9, 10, or another value.

The randomization probability describes the probability of each assignment, not the exact realized allocation.

A More Complete Trial Example

Suppose the trial plans to enroll 100 patients. The investigators use:

• 20-patient burn-in period
• Independent Beta(1,1) priors
• Posterior probability of superiority as the adaptive quantity
• Minimum allocation probability of 20%
• Maximum allocation probability of 80%
• Adaptive allocation beginning after the burn-in period

The basic algorithm becomes:

What Happens If the Experimental Treatment Is Better?

Suppose the true response probabilities are:

As data accumulate, the posterior distribution should increasingly favor larger values of \(p_E\).

Consequently:

will tend to increase over time.

The allocation algorithm may therefore assign an increasing proportion of patients to E.

A stylized sequence might look like:

These numbers are illustrative rather than a guaranteed trajectory. The actual sequence depends on the observed outcomes.

What If the Treatments Are Equal?

Suppose instead:

The posterior should, on average, remain relatively symmetric between the two treatments.

The allocation probabilities should therefore remain closer to balanced allocation.

However, random variation can temporarily create an apparent difference.

The Feedback Problem

This is one of the most important statistical issues in response-adaptive randomization.

Suppose Treatment E receives favorable early outcomes by chance. Then:

The resulting imbalance can affect the information collected from each arm.

This does not automatically invalidate the design. But it means the allocation algorithm should be evaluated through simulation rather than judged solely from its intuitive appeal.

Why Simulation Is Essential

A Bayesian adaptive randomization design is a dynamic system. The allocation probabilities depend on the observed data, while the observed data depend partly on the allocation probabilities.

Therefore, the operating characteristics generally cannot be understood fully from a simple fixed-sample-size formula.

Simulation can estimate:

• Average number of patients assigned to each treatment
• Probability of selecting the better treatment
• Posterior decision probabilities
• Bias of treatment-effect estimates
• Precision of treatment-effect estimates
• Probability of extreme allocation imbalance
• Type I error under the null
• Power under relevant alternatives
• Probability of stopping for efficacy or futility

Frequentist Operating Characteristics Still Matter

A Bayesian design does not mean that frequentist operating characteristics become irrelevant.

For a confirmatory or regulated clinical trial, investigators may still want to evaluate:

• Type I error
• Power
• Coverage probability
• Bias
• Mean squared error
• Allocation imbalance
• Probability of selecting the superior treatment

These properties can be estimated through repeated simulation under specified data-generating mechanisms.

Simulation Setup

Suppose we want to evaluate the design under:

We might simulate 10,000 complete trials. For each simulated trial:

1. Generate the first 20 assignments using 1:1 randomization.
2. Generate the corresponding patient outcomes.
3. Update the posterior distributions.
4. Calculate the adaptive allocation probability.
5. Randomize the next patient.
6. Generate the next outcome.
7. Continue until 100 patients have been enrolled.

R Implementation: Bayesian Updating

# Prior parameters
a_E <- 1
b_E <- 1

a_C <- 1
b_C <- 1

# Observed data
x_E <- 7
n_E <- 10

x_C <- 4
n_C <- 10

# Posterior parameters
a_E_post <- a_E + x_E
b_E_post <- b_E + n_E - x_E

a_C_post <- a_C + x_C
b_C_post <- b_C + n_C - x_C

a_E_post
b_E_post

a_C_post
b_C_post

The resulting posterior distributions are:

# Experimental
# Beta(8, 4)

# Control
# Beta(5, 7)

R Implementation: Posterior Probability of Superiority

set.seed(123)

M <- 1000000

p_E <- rbeta(
  M,
  a_E_post,
  b_E_post
)

p_C <- rbeta(
  M,
  a_C_post,
  b_C_post
)

prob_superior <- mean(
  p_E > p_C
)

prob_superior

The estimated posterior probability is:

where \(I(\cdot)\) is an indicator function.

R Implementation: Bounded Allocation

q <- prob_superior

min_alloc <- 0.20
max_alloc <- 0.80

p_E_next <- min(
  max_alloc,
  max(min_alloc, q)
)

p_C_next <- 1 - p_E_next

p_E_next
p_C_next

This implements:

R Implementation: One Adaptive Trial

simulate_trial <- function(
  N = 100,
  burn_in = 20,
  p_E_true = 0.60,
  p_C_true = 0.40,
  min_alloc = 0.20,
  max_alloc = 0.80
) {

  x_E <- 0
  n_E <- 0

  x_C <- 0
  n_C <- 0

  allocations <- character(N)

  outcomes <- numeric(N)

  for (i in 1:N) {

    if (i <= burn_in) {

      p_E_alloc <- 0.50

    } else {

      M <- 10000

      post_E <- rbeta(
        M,
        1 + x_E,
        1 + n_E - x_E
      )

      post_C <- rbeta(
        M,
        1 + x_C,
        1 + n_C - x_C
      )

      q <- mean(post_E > post_C)

      p_E_alloc <- min(
        max_alloc,
        max(min_alloc, q)
      )
    }

    if (runif(1) < p_E_alloc) {

      allocations[i] <- "E"

      outcome <- rbinom(
        1,
        size = 1,
        prob = p_E_true
      )

      outcomes[i] <- outcome

      x_E <- x_E + outcome
      n_E <- n_E + 1

    } else {

      allocations[i] <- "C"

      outcome <- rbinom(
        1,
        size = 1,
        prob = p_C_true
      )

      outcomes[i] <- outcome

      x_C <- x_C + outcome
      n_C <- n_C + 1
    }
  }

  list(
    allocations = allocations,
    outcomes = outcomes,
    n_E = n_E,
    n_C = n_C,
    x_E = x_E,
    x_C = x_C
  )
}

Run One Simulated Trial

set.seed(2026)

trial <- simulate_trial(
  N = 100,
  burn_in = 20,
  p_E_true = 0.60,
  p_C_true = 0.40
)

trial$n_E
trial$n_C

trial$x_E
trial$x_C

The resulting allocation counts will vary from simulation to simulation. That variability is an intrinsic part of the adaptive randomization process.

Simulating Many Trials

set.seed(2026)

B <- 10000

results <- data.frame(
  n_E = numeric(B),
  n_C = numeric(B),
  x_E = numeric(B),
  x_C = numeric(B)
)

for (b in 1:B) {

  trial <- simulate_trial(
    N = 100,
    burn_in = 20,
    p_E_true = 0.60,
    p_C_true = 0.40
  )

  results$n_E[b] <- trial$n_E
  results$n_C[b] <- trial$n_C

  results$x_E[b] <- trial$x_E
  results$x_C[b] <- trial$x_C
}

Average Allocation

mean(results$n_E)
mean(results$n_C)

These quantities estimate the expected number of patients assigned to each treatment under the assumed true response probabilities.

For a successful adaptive algorithm favoring the superior treatment, the average allocation to E may exceed 50%.

But the magnitude of that increase should be evaluated empirically rather than assumed.

Probability of Correct Treatment Selection

Suppose the experimental treatment is truly superior. A useful operating characteristic is:

For a posterior decision rule such as:

we can estimate the probability of meeting that criterion across simulated trials.

Decision Rules Are Separate From Allocation Rules

This distinction is extremely important. The allocation rule answers:

"How should the next patient be randomized?"

The final decision rule answers:

"What conclusion should be reached at the end of the trial?"

For example, a trial could use:

for allocation, while using:

as the final success criterion.

These are two different components of the design.

Bayesian Adaptive Randomization With Multiple Arms

The concept extends naturally to more than two treatments. Suppose there are four treatment arms:

The design may maintain posterior distributions for:

and then calculate a posterior probability or utility for each arm.

For example:

A softmax allocation rule could then be:

where \(K\) is the number of treatment arms.

Response-Adaptive Randomization vs. Covariate-Adaptive Randomization

The word "adaptive" can refer to different concepts.

These approaches should not be treated as interchangeable.

Bayesian Adaptive Randomization vs. Stratified Randomization

Suppose a conventional trial uses stratified randomization by disease stage. The randomization probability might depend on whether a patient is:

• Stage II
• Stage III
• Stage IV

That is a form of covariate control.

Bayesian response-adaptive randomization instead changes allocation according to accumulating treatment-response information.

The two approaches can also be combined.

Delayed Outcomes Create an Important Challenge

Many clinical endpoints are not observed immediately. For example, suppose the primary endpoint requires 12 weeks of follow-up.

Patient 51 may be randomized today while patient 31's primary endpoint is still pending.

The adaptive algorithm must specify how pending outcomes are handled.

Possible approaches include:

• Updating only when mature outcomes become available
• Using interim or short-term outcomes as predictive information
• Using models that account for partially observed follow-up
• Using delayed-response adaptive randomization methods

Patient-Level vs. Batch Adaptation

Adaptation does not necessarily need to occur after every patient.

The design might instead update after batches of patients.

For example:

Batch adaptation can be operationally easier and can prevent allocation probabilities from changing excessively frequently.

Why the Timing of Adaptation Matters

If allocation probabilities change after every single outcome, the algorithm can react strongly to random fluctuations.

If probabilities are updated only occasionally, the algorithm may be more stable but less responsive.

Therefore, the design must specify:

• When posterior updates occur
• Which outcomes are considered mature
• How pending outcomes are handled
• When new allocation probabilities become active

Prior Distributions Matter

Bayesian adaptive randomization depends on the prior distribution. Suppose:

Both priors have mean 0.50. But they express different amounts of prior information.

For Beta(1,1):

whereas for Beta(10,10):

The second prior is substantially more concentrated around 0.50.

Consequently, the same early clinical data can produce different posterior distributions under the two priors.

Prior Effective Sample Size

For a Beta prior:

a useful heuristic for prior effective sample size is:

For Beta(1,1):

For Beta(10,10):

This illustrates why apparently similar priors can have very different effects during the early stages of a trial.

Robustness to the Prior

A strong design analysis should consider multiple plausible priors.

For example:

The goal is to determine whether the adaptive allocation behavior is robust to reasonable changes in prior assumptions.

Historical Control Information

Bayesian models can incorporate historical information about control response. For example:

where the parameters are chosen to represent external information.

This can reduce the amount of concurrent control information required for certain inferential objectives.

However, historical information should be used cautiously because differences in:

• Patient population
• Eligibility criteria
• Standard of care
• Endpoint definition
• Assessment timing
• Study conduct

can make historical data nonexchangeable with the current trial.

Exchangeability Is a Modeling Assumption

Suppose historical control patients are represented by \(H\) and current control patients by \(C\). Pooling them implicitly assumes that their response probabilities are sufficiently comparable.

If that assumption is wrong, the posterior can become overly confident.

More sophisticated Bayesian models can introduce an explicit discounting or commensurability mechanism rather than treating historical and current data as perfectly exchangeable.

Safety Should Not Be Ignored

A response-adaptive algorithm based only on efficacy can allocate more patients to a treatment that appears effective but has an important safety problem.

Therefore, a clinically realistic adaptive design may incorporate both:

• Efficacy
• Safety

For example, an allocation utility might depend on:

where \(w_E\) and \(w_S\) reflect prespecified utility weights.

The exact utility model depends on the clinical context.

Bayesian Adaptive Randomization Does Not Mean "Always Give the Best Treatment"

This is one of the most common misconceptions.

The algorithm does not know which treatment is truly best. It only has a posterior distribution describing uncertainty about the unknown parameters.

If:

the model is not saying that E is definitely superior. It is saying that, under the model and prior, there is 70% posterior probability that E is superior.

The remaining uncertainty is precisely why randomization continues.

Common Mistakes

1. Confusing posterior probability with a p-value. A posterior probability such as \(P(p_E>p_C\mid D)\) is not a frequentist p-value.
2. Assuming the allocation probability must equal posterior probability. The posterior probability is an input to the adaptive rule, not the rule itself.
3. Ignoring the prior. Bayesian adaptive randomization depends on the specified prior distributions.
4. Allowing allocation probabilities to become arbitrarily extreme. This can create severe imbalance and reduce information about the comparison.
5. Ignoring delayed outcomes. The adaptive algorithm must specify how pending primary endpoints are handled.
6. Evaluating the design only through Bayesian posterior quantities. Simulation-based frequentist operating characteristics can still be essential.
7. Changing the allocation rule after observing results. The adaptation algorithm should be prespecified.
8. Ignoring treatment-by-covariate interactions. An allocation strategy based on a marginal treatment effect may not perform well when treatment effects vary substantially across patient subgroups.
9. Assuming adaptive randomization automatically improves the trial. Adaptive allocation involves statistical and operational tradeoffs that need to be evaluated for the specific study.
10. Using an adaptive algorithm without simulation. Because allocation and outcomes interact dynamically, simulation is generally necessary to understand operating characteristics.

Bayesian Adaptive Randomization Workflow

Operating Characteristics to Evaluate

A Bayesian adaptive randomization design should be evaluated across a broad range of scenarios.

Scenario-Based Simulation

A useful simulation program evaluates multiple true treatment-response scenarios.

This type of scenario analysis is substantially more informative than evaluating the algorithm under only one assumed treatment effect.

Bayesian Adaptive Randomization and Equipoise

Traditional clinical trial randomization is often justified partly by maintaining uncertainty about which treatment is superior.

Adaptive randomization changes the allocation as that uncertainty changes.

This creates a fundamental design tension:

A sound design therefore needs to define how much adaptation is appropriate for its scientific objective.

Ethical Motivation

One potential motivation for adaptive randomization is that patients enrolled later in a trial may have a greater probability of receiving a treatment that currently appears promising.

However, this potential ethical benefit should be evaluated alongside the statistical consequences of unequal allocation.

If adaptation becomes too aggressive, the trial may obtain less information about the comparison or become more sensitive to early random variation.

Adaptive Randomization in Platform Trials

Bayesian adaptive randomization can be especially relevant in multi-arm or platform trials.

Suppose a platform contains:

• Shared control
• Experimental A
• Experimental B
• Experimental C

Posterior probabilities can be updated for all active treatments.

An adaptive allocation mechanism can then shift enrollment toward arms with greater posterior utility.

At the same time, arms can be dropped if prespecified futility criteria are met.

This creates a broader adaptive ecosystem involving:

• Response-adaptive randomization
• Arm dropping
• Arm addition
• Sample-size adaptation
• Bayesian predictive probabilities
• Shared control groups

Bayesian Adaptive Randomization and MCMC

For the simple Beta-Binomial model, posterior calculations are analytically tractable.

But real clinical-trial models may involve:

• Continuous endpoints
• Time-to-event outcomes
• Hierarchical models
• Longitudinal outcomes
• Patient-level covariates
• Missing data
• Historical information
• Nonconjugate priors

In these settings, the posterior distribution may not have a closed form.

Markov Chain Monte Carlo can then be used to generate posterior samples:

Posterior probabilities and expected utilities can then be estimated from those samples.

For example:

This connects Bayesian adaptive randomization directly to MCMC-based Bayesian clinical-trial modeling.

Posterior Predictive Probability

Another adaptive quantity is the posterior predictive probability of eventual trial success.

Suppose the final success criterion is:

The posterior predictive probability might be:

This asks a different question from posterior probability of superiority.

Instead of asking:

"Is E currently better than C?"

it asks:

"Given what we know now, how likely is the final trial to satisfy the success criterion?"

This distinction becomes important when adaptive randomization is combined with Bayesian interim monitoring.

Allocation vs. Monitoring

These components can coexist, but each should have a clearly defined role.

Protocol Considerations

The adaptive randomization algorithm should be fully described in the protocol or associated statistical documentation.

At minimum, specify:

• Primary endpoint
• Bayesian likelihood
• Prior distributions
• Parameterization of treatment effects
• Burn-in period
• Frequency of posterior updating
• Definition of outcome maturity
• Adaptive allocation function
• Minimum allocation probability
• Maximum allocation probability
• Handling of delayed outcomes
• Handling of missing outcomes
• Safety rules
• Futility rules
• Efficacy rules
• Final decision criterion
• Simulation scenarios used for design evaluation

Operational Implementation

In a real clinical trial, allocation probabilities are generally generated by a controlled randomization system rather than calculated manually by study personnel.

The operational system needs to ensure:

• Correct implementation of the algorithm
• Secure treatment assignment
• Preservation of blinding where applicable
• Accurate transfer of response data
• Correct timing of posterior updates
• Auditability of allocation changes
• Version control for the adaptive algorithm

Validation of the Randomization Algorithm

Before deployment, the implementation should be tested against known scenarios.

For example, if the posterior probability is exactly:

the allocation algorithm should produce:

If:

with an 80% maximum allocation cap, the implementation should return:

These deterministic test cases should be supplemented with simulation-based validation.

Common Misinterpretation: "Adaptive Means Unblinded"

An adaptive design does not necessarily mean that investigators or clinical site personnel see accumulating comparative results.

In a properly controlled trial, the adaptive algorithm can operate centrally using data that are not visible to the clinical team.

The exact operational structure depends on the trial design and governance arrangements.

Common Misinterpretation: "Bayesian Means No Type I Error"

Bayesian designs do not eliminate false-positive decisions.

If a final decision is made using a posterior probability threshold such as:

the probability of declaring success when the treatment is actually ineffective still depends on the prior, model, sample size, stopping rules, and data generating mechanism.

That is why operating characteristics should be evaluated through simulation under relevant null scenarios.

A Practical Design Template

A simple two-arm Bayesian adaptive randomization design might be specified as:

At each adaptive update:

Then:

and:

This algorithm continues until a prespecified stopping criterion or maximum sample size is reached.

Worked Example Summary

The Most Important Concept

The central idea behind Bayesian adaptive randomization is simple:

The trial continuously moves through this feedback loop.

Observed outcomes update the posterior distribution. The posterior distribution determines the current evidence about treatment performance. That evidence is converted into a prespecified allocation probability. The next patient is randomized according to that probability.

The process then repeats.

The statistical challenge is not simply constructing a Bayesian posterior. It is constructing an adaptive system whose behavior remains scientifically, statistically, ethically, and operationally defensible under a wide range of possible realities.

References

Berry, D.A. (2006). Bayesian clinical trials. Nature Reviews Drug Discovery, 5, 27–36.
Berry, S.M., Carlin, B.P., Lee, J.J. & Muller, P. (2010). Bayesian Adaptive Methods for Clinical Trials. CRC Press.
Hu, F. & Rosenberger, W.F. (2006). The Theory of Response-Adaptive Randomization in Clinical Trials. Wiley.
Rosenberger, W.F. & Lachin, J.M. (2016). Randomization in Clinical Trials: Theory and Practice. Wiley.
Thall, P.F. & Wathen, J.K. (2007). Practical Bayesian adaptive randomisation in clinical trials. European Journal of Cancer, 43(5), 859–866.
Berry, S.M., Connor, J.T. & Lewis, R.J. (2015). The platform trial: an efficient strategy for evaluating multiple treatments. JAMA, 313(16), 1619–1620.
Viele, K., Berry, S., Neuenschwander, B., et al. (2014). Use of historical control data for assessing treatment effects in clinical trials. Pharmaceutical Statistics, 13(1), 41–54.