Adaptive Randomization Methods

Introduction

Randomization is one of the central principles of controlled clinical trials. Its purpose is to create treatment groups that are comparable with respect to both measured and unmeasured prognostic factors.

The simplest approach is fixed-probability randomization. For example, in a two-arm trial with equal allocation, every eligible patient has a 50% probability of receiving Treatment A and a 50% probability of receiving Treatment B.

Adaptive randomization changes this framework. Instead of requiring the same allocation probability for every patient throughout the trial, the probability of assignment may depend on information available at the time of randomization.

Fixed Randomization vs. Adaptive Randomization

Suppose a two-arm trial uses equal allocation. Under conventional randomization:

for every patient.

Under an adaptive procedure, the allocation probabilities may instead be:

where \(i\) denotes the incoming patient and \(\pi_i\) may depend on information accumulated before patient \(i\) is randomized.

For example:

The changing probabilities are what make the procedure adaptive.

Why Would We Adapt Randomization?

There are several fundamentally different reasons for changing allocation probabilities.

The first is to improve balance across prognostic covariates. For example, investigators may want similar distributions of age, disease severity, or biomarker status across treatment groups.

The second is to use accumulating outcome information to allocate more patients toward treatments that appear to perform better.

The third is to incorporate both covariates and outcomes into a Bayesian or model-based decision process.

Three Major Categories

A useful conceptual classification is:

Covariate-Adaptive Randomization

Suppose a randomized trial compares Treatments A and B and the investigators are particularly concerned about balancing:

• Age category
• Sex
• Disease stage
• Baseline biomarker status

Simple randomization will usually produce reasonable balance in a sufficiently large trial, but meaningful imbalances can occur by chance, particularly in small studies.

Covariate-adaptive randomization attempts to reduce those imbalances during enrollment.

Stratified Randomization

One familiar approach is to define strata based on important baseline covariates.

For example:

Within each stratum, a separate randomization sequence can be used.

The advantage is straightforward: important covariates are controlled directly during treatment assignment.

The disadvantage is that the number of strata can grow rapidly as additional categorical covariates are included.

where \(L_j\) is the number of levels of covariate \(j\).

For four binary covariates, this already gives:

strata.

What Is Minimization?

Minimization is a covariate-adaptive allocation method designed to balance multiple prognostic factors simultaneously.

Rather than creating a separate randomization list for every possible combination of covariates, the procedure evaluates how assigning the incoming patient to each treatment would affect overall imbalance.

Suppose a new patient has:

• Age ≥65
• Male sex
• Stage III disease

The algorithm calculates the imbalance that would result from assigning the patient to Treatment A and separately calculates the imbalance that would result from assigning the patient to Treatment B.

The treatment that produces the smaller imbalance receives the larger allocation probability.

A Simple Minimization Score

Let:

be the imbalance score if the incoming patient is assigned to Treatment A, and:

the corresponding score for Treatment B.

A simple deterministic rule would assign the patient to the treatment with the smaller score.

However, deterministic minimization makes the treatment assignment increasingly predictable.

A common approach is therefore to retain a random component.

where \(q\) might be 0.80 or 0.90.

For example, with \(q=0.80\), the treatment that produces better balance has an 80% probability of being selected, while the other treatment retains a 20% probability.

Worked Minimization Example

Suppose a trial has two treatment groups and uses three binary covariates:

• Age: <65 vs. ≥65
• Sex: female vs. male
• Disease severity: low vs. high

Before the next patient is randomized, the current counts are:

The incoming patient is:

• Age ≥65
• Male
• High severity

If assigned to Drug A, the corresponding marginal differences would become:

If assigned to Drug B:

The overall imbalance score can then be defined using a prespecified function, such as a weighted sum of absolute differences.

and:

If \(I_B

Response-Adaptive Randomization

Response-adaptive randomization changes treatment allocation according to observed outcomes.

Suppose a trial begins with equal allocation:

After accumulating outcome information, Treatment A appears more effective. The allocation probabilities might then change to:

Later, if evidence becomes stronger:

The intention is to expose more future patients to the treatment that appears more beneficial.

The Allocation Probability Function

Let:

represent treatment-specific efficacy parameters. A response-adaptive algorithm can define the probability of assigning the next patient to Treatment A as a function:

and:

The function \(f(\cdot)\) determines how aggressively the allocation responds to accumulating evidence.

Why Not Always Allocate to the Better Treatment?

A natural idea is: if Treatment A looks better, assign the next patient to A.

This is generally too aggressive.

Early in a trial, the estimated treatment effect can be highly unstable.

For example, suppose after only 10 patients:

It would be premature to conclude that A truly has an 80% response rate and B a 40% response rate.

Adaptive procedures therefore usually include mechanisms that prevent small random fluctuations from producing extreme allocation probabilities.

Randomized-Urn and Play-the-Winner Designs

Some adaptive procedures use urn models.

Conceptually, an urn contains treatment labels. When a treatment produces a favorable outcome, the procedure may add additional labels favoring that treatment.

As favorable outcomes accumulate, the probability of assigning subsequent patients to that treatment increases.

These approaches are related to the broader family of play-the-winner designs.

Bayesian Adaptive Randomization

Bayesian adaptive randomization provides a particularly intuitive framework for incorporating accumulating evidence.

Suppose Treatment A has response probability:

and Treatment B has response probability:

After observing data, a Bayesian analysis produces posterior distributions:

The randomization probability can then depend on the posterior probability that one treatment is better than another.

Posterior Probability of Being Best

Suppose simulation from the posterior distributions gives:

A Bayesian adaptive procedure might transform this posterior probability into an allocation probability.

For example, a simple rule might be:

although in practice the relationship may be softened or modified to avoid overly aggressive allocation.

A Bayesian Beta-Binomial Example

Suppose binary responses are observed and each treatment has a Beta prior:

Suppose after an interim analysis:

The posterior distributions are:

and:

The posterior means are:

The posterior probability that A is better than B can be calculated by integrating over the joint posterior distribution:

where \(f_A\) is the posterior density of \(p_A\) and \(F_B\) is the posterior CDF of \(p_B\).

This probability can be estimated efficiently using posterior simulation.

Posterior Simulation

set.seed(123)

draws <- 100000

pA <- rbeta(
  draws,
  19,
  13
)

pB <- rbeta(
  draws,
  13,
  19
)

prob_A_better <- mean(
  pA > pB
)

prob_A_better

The resulting value estimates:

This posterior probability can then be used by the adaptive allocation algorithm.

Randomization Should Usually Remain Random

A key design principle is that adaptive randomization should not generally collapse into deterministic treatment assignment.

One simple way to control this is to impose probability bounds:

For example, with:

the allocation probability is restricted to:

Even if Treatment A appears substantially superior, at least 20% of future patients remain eligible for randomization to Treatment B.

Response-Adaptive Allocation Functions

There are many ways to transform estimated treatment performance into an allocation probability.

A simple power transformation is:

where \(\gamma\) controls the aggressiveness of adaptation.

If:

then:

regardless of estimated response rates.

As \(\gamma\) increases, the allocation becomes increasingly concentrated on the treatment with the larger estimated response rate.

Worked Allocation Example

Suppose:

and choose:

Then:

which gives:

and:

Thus approximately 69% of future patients would be assigned to A and 31% to B under this particular allocation rule.

Why the Algorithm Matters More Than the Observed Allocation

An adaptive trial may ultimately produce an allocation ratio such as:

That does not by itself tell us whether the design was valid.

The statistical properties depend on how that allocation ratio was generated.

A prespecified adaptive algorithm might intentionally produce a 70:30 ratio. A biased enrollment process could also accidentally produce a 70:30 ratio. The observed allocation alone does not distinguish the two.

Randomization Probability vs. Observed Allocation Ratio

These concepts should be distinguished carefully.

The randomization probability is:

where \(D_{i-1}\) represents information available before patient \(i\) is randomized.

The observed allocation fraction is:

These quantities are related but not identical.

Even with a fixed 50% randomization probability, the observed allocation fraction will not necessarily be exactly 50%.

Adaptation Can Increase Variability

Adaptive randomization introduces another layer of stochastic dependence. Under fixed randomization:

is often a useful conceptual representation of independent treatment assignments under a fixed allocation probability.

Under response-adaptive randomization, the probability for patient \(i\) depends on previous observations:

and therefore the treatment assignments are no longer identically distributed in the simple fixed-probability sense.

This is one reason ordinary textbook sample-size and variance formulas may not adequately describe an adaptive trial.

The Role of Simulation

Simulation is one of the most important tools for evaluating adaptive randomization.

A simulation repeatedly generates complete trials under specified assumptions and applies the exact adaptive algorithm.

Type I Error in Adaptive Randomization

Suppose the null hypothesis is:

The type I error is:

The adaptive allocation procedure itself does not automatically guarantee a particular type I error.

The complete design—including allocation algorithm, interim analyses, stopping rules, final analysis, and adaptation schedule—must be evaluated.

Power Under Adaptive Randomization

Power remains:

However, power can change when the allocation mechanism changes.

For example, if an adaptive procedure increasingly favors Treatment A, fewer patients may receive Treatment B.

That can be beneficial if A is truly superior and the scientific objective is to expose more patients to the better treatment.

But it may also reduce the amount of information available for estimating the difference between A and B.

The Exploration-Exploitation Tradeoff

Adaptive randomization can be understood as balancing two competing objectives.

Exploration

Continue assigning patients to all treatments so that the trial learns which treatments are effective.

Exploitation

Allocate more patients to treatments that currently appear superior.

If the design exploits too aggressively, early random variation can cause the trial to favor the wrong treatment.

If it explores too aggressively, the trial may provide little advantage over conventional equal randomization.

Why Early Random Variation Is Dangerous

Suppose the true response rates are:

The true difference is only:

If the first few patients happen to produce more responses on B, an aggressive adaptive algorithm could incorrectly shift allocation toward B.

That creates a feedback loop:

This phenomenon is one reason adaptive randomization must be carefully calibrated.

Response-Adaptive Randomization Can Be Ethically Attractive

The ethical argument for response-adaptive randomization is intuitive. If accumulating evidence suggests that one treatment is substantially better, future participants may have a higher probability of receiving that treatment.

For example:

means that three out of four future patients, on average, will receive A.

This can reduce expected exposure to an apparently inferior treatment.

Inference After Adaptive Randomization

One of the most important practical questions is: How should the treatment effect be analyzed after adaptive randomization?

The answer depends on the allocation mechanism.

For simple covariate-adaptive procedures, standard regression methods may often be appropriate when the analysis properly accounts for the design and relevant covariates.

For response-adaptive procedures, however, the treatment assignment mechanism can depend on previous outcomes.

This creates additional statistical complexity.

Why Naive Analysis Can Be Misleading

Suppose a trial progressively allocates more patients to a treatment as that treatment appears better.

The observed treatment groups are no longer generated by a simple fixed allocation mechanism.

If the analysis ignores the adaptive mechanism, standard estimates and standard errors may not have the operating characteristics assumed by a conventional randomized trial.

Therefore, the allocation rule and analysis model should be designed together.

Covariate Adjustment Can Improve Precision

Suppose the primary endpoint is continuous:

where:

• \(A_i\) is the treatment indicator
• \(X_i\) is a vector of baseline covariates
• \(\tau\) is the treatment effect

Even when randomization balances covariates reasonably well, covariate adjustment can improve precision.

This is an important distinction: adaptive randomization and covariate-adjusted analysis are not alternatives. They can be used together.

Dynamic Allocation

The general concept behind minimization and related procedures is sometimes called dynamic allocation.

Instead of generating one fixed randomization sequence before the trial begins, the treatment assignment is calculated dynamically for each incoming patient.

where \(D_{i-1}\) represents information accumulated before patient \(i\) and \(X_i\) represents the incoming patient's baseline covariates.

This allows the allocation to respond immediately to the current composition of the trial.

Covariate-Adaptive vs. Response-Adaptive Randomization

Bayesian Adaptive Randomization vs. Response-Adaptive Randomization

These terms are sometimes used interchangeably, but they need not describe the same procedure.

Response-adaptive randomization is defined by the use of accumulating response information.

Bayesian adaptive randomization describes the statistical framework used to update beliefs about treatment effects.

A Bayesian procedure can therefore be response-adaptive, but Bayesian inference and adaptive allocation are conceptually separate components.

Example: Bayesian Allocation Rule

Suppose posterior simulation gives:

and:

A simple probability-matching strategy sets:

A tempered strategy might instead use:

with \(0<\gamma<1\) to reduce the degree of adaptation.

For example, with:

the allocation probabilities move closer to 50:50.

Delayed Outcomes Create a Major Challenge

Response-adaptive randomization works most naturally when outcomes become available quickly.

Suppose the primary endpoint requires six months of follow-up.

The trial might randomize hundreds of patients before the first patients' outcomes become available.

In that setting, a response-adaptive algorithm cannot respond rapidly to accumulating information.

This can motivate:

• Delayed-outcome models
• Time-to-event adaptive methods
• Interim analyses based on mature information
• Joint models for incomplete outcomes

Time-to-Event Outcomes

For survival endpoints, the relevant treatment parameter might be a hazard ratio:

Adaptive allocation could depend on posterior or estimated information about the hazard ratio.

However, survival data introduce additional complications because patients have different follow-up durations and censoring patterns.

The allocation model therefore needs to account for incomplete information rather than treating every enrolled patient as having a fully observed outcome.

Multi-Arm Trials

Adaptive randomization becomes particularly interesting when a trial has more than two treatments.

Suppose there are four experimental treatments and one control:

A fixed equal-allocation design assigns approximately:

to each arm.

An adaptive procedure might eventually produce:

Treatment C receives more patients because its accumulating evidence is more favorable.

This can reduce exposure to inferior experimental arms.

Maintaining a Control Group

In many adaptive multi-arm trials, investigators impose a minimum allocation to control.

For example:

regardless of how favorable the experimental arms appear.

This preserves a sufficiently large contemporaneous control group for interpretable comparisons.

Allocation Constraints

A general constrained adaptive allocation problem can be written:

for each treatment \(j\), subject to:

These constraints prevent any treatment from receiving essentially all or none of the randomizations.

What Does "Adaptive" Mean Statistically?

The essential feature is conditional dependence.

The probability of assigning patient \(i\) to treatment \(j\) can depend on information available before that assignment:

Under fixed randomization:

for every patient.

Under adaptive randomization:

so the probability changes according to available information.

Allocation Concealment

Adaptive randomization creates special challenges for allocation concealment.

If investigators know:

• The current treatment counts
• The current allocation probabilities
• The exact algorithm
• The covariates of the incoming patient

they may be able to predict the next assignment with substantial accuracy.

Therefore, adaptive randomization should generally be implemented through a centralized system that calculates the assignment automatically.

Predictability Is a Design Property

A useful way to evaluate an allocation method is to ask: How predictable is the next assignment?

For a two-arm design with:

the next treatment is maximally unpredictable.

If:

the assignment is nearly deterministic.

Highly predictable allocation can create operational and methodological problems even if the algorithm has desirable theoretical properties.

Randomization Tests

Because adaptive randomization generates treatment assignments according to a known probability mechanism, randomization-based inference may sometimes be considered.

The fundamental idea is to evaluate the observed statistic against the distribution generated by the randomization mechanism under the null.

Conceptually:

This emphasizes an important point: the randomization rule is part of the statistical design.

Simulation-Based Operating Characteristics

For a complex adaptive design, closed-form calculations may be unavailable. Simulation can estimate:

A Basic Simulation Framework in R

set.seed(123)

nsim <- 5000
n <- 200

true_p_A <- 0.50
true_p_B <- 0.35

results <- matrix(
  NA,
  nrow = nsim,
  ncol = 3
)

colnames(results) <- c(
  "n_A",
  "n_B",
  "effect"
)

A simple adaptive allocation function might be defined as:

allocation_prob <- function(
  successes_A,
  n_A,
  successes_B,
  n_B
) {

  rate_A <- (
    successes_A + 1
  ) / (
    n_A + 2
  )

  rate_B <- (
    successes_B + 1
  ) / (
    n_B + 2
  )

  raw <- rate_A /
    (rate_A + rate_B)

  max(
    0.20,
    min(0.80, raw)
  )
}

This example uses a simple smoothed response-rate estimate and constrains the allocation probability to the interval 0.20 to 0.80.

A complete simulation could then proceed sequentially:

for (s in 1:nsim) {

  n_A <- 0
  n_B <- 0

  y_A <- 0
  y_B <- 0

  for (i in 1:n) {

    p_A <- allocation_prob(
      y_A,
      n_A,
      y_B,
      n_B
    )

    assign_A <- runif(1) < p_A

    if (assign_A) {

      y <- rbinom(
        1,
        1,
        true_p_A
      )

      n_A <- n_A + 1
      y_A <- y_A + y

    } else {

      y <- rbinom(
        1,
        1,
        true_p_B
      )

      n_B <- n_B + 1
      y_B <- y_B + y
    }
  }

  results[s, "n_A"] <- n_A
  results[s, "n_B"] <- n_B

  results[s, "effect"] <-
    y_A / n_A -
    y_B / n_B
}

This is deliberately simplified, but it demonstrates the central structure of adaptive randomization:

Estimating the Allocation Distribution

After simulation, investigators can examine the distribution of final allocation fractions.

mean(
  results[, "n_A"] / n
)

mean(
  results[, "n_B"] / n
)

They can also calculate quantiles:

quantile(
  results[, "n_A"] / n,
  c(0.05, 0.50, 0.95)
)

This reveals how much the allocation ratio varies across simulated trials.

Evaluating Probability of Selecting the Best Treatment

Suppose Treatment A is truly better:

A useful operating characteristic is the probability that the adaptive procedure allocates the majority of patients to A:

More sophisticated designs may instead evaluate:

at the final decision.

Null Scenarios Are Essential

Adaptive algorithms should not be evaluated only when one treatment is better.

At minimum, simulation should usually consider:

• Equal treatment efficacy
• Small treatment differences
• Moderate treatment differences
• Large treatment differences
• Different baseline event rates
• Potential model misspecification

The equal-efficacy scenario is especially important because it can reveal whether the adaptive algorithm creates systematic allocation imbalance even when no treatment is truly superior.

Example Scenario Grid

Adaptive Randomization and Sample Size

Adaptive allocation can affect the information available for treatment comparisons.

For a two-arm trial with approximately equal allocation, the variance of the difference between two independent proportions is approximately:

For a fixed total sample size:

extreme imbalance can increase the variance of comparative estimates.

This is one reason adaptive allocation can create a tension between patient benefit and statistical efficiency.

Why 1:1 Allocation Is Often Statistically Efficient

When the two treatment groups have similar outcome variance and the primary objective is estimating a treatment contrast, approximately equal allocation often provides high statistical efficiency.

For a simple two-group comparison, the variance is minimized near:

Therefore, deliberately moving away from 1:1 allocation has a statistical cost.

Adaptive randomization must provide a sufficiently important benefit—such as ethical allocation or covariate balance—to justify that cost.

When Unequal Allocation Is Already Appropriate

Unequal allocation is not inherently adaptive.

For example, a trial might deliberately use:

randomization so that more patients receive the experimental treatment.

That is a fixed unequal allocation design, not adaptive randomization.

Adaptive randomization requires that the allocation probability can change in response to information.

Adaptive Randomization Does Not Automatically Reduce Sample Size

Another common misconception is that adaptive allocation necessarily makes trials smaller.

It may improve ethical allocation, but the total sample size required to achieve a particular statistical objective may be unchanged or even larger.

Sample size depends on:

• Target effect size
• Outcome variance
• Type I error
• Power
• Allocation ratio
• Interim analyses
• Adaptation rule
• Stopping rules

Adaptive Randomization vs. Adaptive Sample Size Re-Estimation

These are separate concepts.

Adaptive Randomization vs. Group Sequential Design

These concepts are also distinct.

A group sequential design changes the decision to stop or continue the trial at interim analyses.

Adaptive randomization changes the probability of assigning patients to treatments.

The two can be combined.

Combining Adaptive Randomization With Early Stopping

A modern adaptive trial may use several adaptations simultaneously.

For example:

• Response-adaptive allocation
• Early futility stopping
• Early efficacy stopping
• Sample size adaptation
• Arm dropping

The complexity of the combined design means that simulation becomes especially important.

Arm Dropping

Suppose a trial begins with four experimental treatments:

An interim analysis may determine that D has little chance of being useful.

The trial can then drop D and redistribute future patients among the remaining arms.

This is an example of adaptive treatment selection and is related to, but distinct from, adaptive randomization.

Adaptive Randomization in Platform Trials

Platform trials are especially well suited to adaptive allocation because multiple treatments can be evaluated simultaneously.

A platform may:

• Add new treatment arms
• Drop ineffective arms
• Change allocation probabilities
• Share a common control
• Adapt according to biomarker-defined subgroups

The statistical design becomes considerably more complex because treatment comparisons may share controls and information across time.

Biomarker-Adaptive Randomization

Suppose a treatment is expected to work particularly well in biomarker-positive patients.

An adaptive design can incorporate biomarker status into allocation.

For example:

This is different from simply adapting based on overall treatment performance. The algorithm can target treatment allocation to the patient population in which the treatment appears most beneficial.

Subgroup Adaptation Requires Care

If treatment effects differ across subgroups, adaptive allocation can produce very different sample sizes within each subgroup.

Investigators should therefore evaluate:

• Precision within each subgroup
• Probability of identifying the correct treatment
• Potential treatment-by-subgroup interaction
• Minimum subgroup sample sizes
• Control-group representation

Advantages of Adaptive Randomization

• Potential ethical benefit: more patients may receive treatments that appear beneficial.
• Covariate balance: prognostic factors can be balanced dynamically.
• Flexibility: allocation can respond to accumulating information.
• Efficient multi-arm exploration: weak treatments may receive fewer patients.
• Personalization: allocation can incorporate patient-level covariates.

Limitations of Adaptive Randomization

• Statistical complexity: inference can be more complicated than under fixed randomization.
• Predictability: aggressive adaptation can make future assignments easier to anticipate.
• Early noise: random fluctuations can influence later allocation.
• Reduced comparative information: extreme imbalance can reduce precision.
• Operational complexity: centralized algorithms and rapid data transfer may be required.
• Model dependence: Bayesian or model-based approaches can be sensitive to assumptions.
• Regulatory complexity: the complete adaptation strategy must be justified and prespecified.

Common Mistakes

1. Calling any unequal allocation adaptive. A fixed 2:1 design is not adaptive simply because it is unequal.
2. Using observed outcomes to choose treatment manually. Adaptive allocation must follow a prespecified rule.
3. Ignoring early random variation. An aggressive response-adaptive algorithm can reinforce random noise.
4. Allowing allocation probabilities to approach zero. Extreme allocation can create weak comparative information.
5. Ignoring delayed outcomes. An adaptive algorithm cannot respond to information that is not yet available.
6. Assuming adaptive randomization automatically improves power. It can sometimes reduce information for the treatment contrast.
7. Failing to simulate under the null. The equal-treatment scenario is essential for evaluating type I error and allocation behavior.
8. Using standard inference without considering the allocation mechanism. Adaptive assignment can alter the properties of conventional estimators and tests.
9. Allowing investigators to predict assignments. Predictability can undermine allocation concealment.
10. Changing the algorithm after observing results. The statistical properties of the original design no longer necessarily apply.

Practical Design Workflow

What Should Be Prespecified?

An adaptive randomization strategy should be sufficiently detailed that an independent statistician could reproduce the assignment mechanism.

The protocol or statistical analysis documentation should specify:

• Number of treatment arms
• Initial allocation probabilities
• Variables used for adaptation
• Definitions of all covariates
• Outcome definitions
• Timing of outcome updates
• Allocation formula
• Probability constraints
• Randomization mechanism
• Frequency of adaptation
• Handling of missing data
• Handling of delayed outcomes
• Primary analysis method
• Interim monitoring procedures
• Stopping rules
• Simulation scenarios and operating characteristics

A Simple Decision Framework

When Fixed Randomization May Be Better

Adaptive randomization is not automatically superior.

Fixed randomization may be preferable when:

• The trial is large enough that covariate imbalance is unlikely to be problematic.
• The primary objective is highly efficient estimation of a treatment contrast.
• There is no compelling ethical reason to favor one treatment.
• Outcome ascertainment is substantially delayed.
• The trial requires simple and transparent statistical inference.
• Operational simplicity is important.
• Allocation concealment is particularly sensitive.

Adaptive Randomization Is Part of the Trial Design

The allocation mechanism should not be viewed as an isolated programming component.

It interacts with:

• Sample size
• Interim monitoring
• Endpoint timing
• Statistical analysis
• Missing-data handling
• Allocation concealment
• Operational logistics

A change to one component can affect the operating characteristics of the others.

Worked Conceptual Example

Consider a randomized Phase II trial comparing A and B with a binary response endpoint. The trial begins with:

After the first 40 evaluable patients, the observed response rates are:

An adaptive algorithm might update the next allocation to:

After another 40 patients, suppose the accumulated results are:

The algorithm might then increase allocation to A again, perhaps to:

This illustrates the basic feedback mechanism.

However, whether this is a good design cannot be determined from these observed rates alone.

The complete procedure must be evaluated through simulation.

What Simulation Should Demonstrate

For the example above, a rigorous simulation study should answer:

• Does the procedure preserve type I error when \(p_A=p_B\)?
• How much does the final allocation favor the superior treatment?
• What happens when A is only slightly better?
• How often does the algorithm initially favor the wrong treatment?
• What is the bias of the treatment-effect estimator?
• What is the standard error?
• How much statistical information is lost through unequal allocation?
• How sensitive are results to delayed outcomes?

The Central Statistical Tradeoff

Adaptive randomization can be summarized as a tradeoff between:

Allocating more patients to an apparently superior treatment can be ethically appealing.

But comparative inference generally benefits from assigning substantial numbers of patients to all treatments.

The best design depends on the scientific objective.

Key Takeaways

Adaptive randomization is a family of prespecified allocation procedures in which treatment assignment probabilities can depend on information available during the trial.

The major classes include:

• Covariate-adaptive randomization for improving baseline balance.
• Minimization for dynamically balancing multiple prognostic factors.
• Response-adaptive randomization for shifting allocation according to accumulating outcomes.
• Bayesian adaptive randomization for using posterior treatment information.
• Play-the-winner and urn methods for modifying future allocation based on observed responses.

The mathematical centerpiece is the conditional allocation probability:

Unlike fixed randomization, the probability can change from one patient to the next.

The design must therefore be evaluated not only for allocation behavior but also for:

• Type I error
• Power
• Bias
• Precision
• Sample size
• Allocation ratios
• Probability of favoring the truly best treatment
• Predictability
• Robustness to model assumptions

References

Taves, D.R. (1974). Minimization: A new method of assigning patients to treatment and control groups. Clinical Pharmacology & Therapeutics, 15(5), 443–453.
Pocock, S.J. & Simon, R. (1975). Sequential treatment assignment with balancing for prognostic factors in the controlled clinical trial. Biometrics, 31(1), 103–115.
Efron, B. (1971). Forcing a sequential experiment to be balanced. Biometrika, 58(3), 403–417.
Wei, L.J. & Durham, S. (1978). The randomized play-the-winner rule in medical trials. Journal of the American Statistical Association, 73(364), 840–843.
Wei, L.J. (1978). An application of an urn model to the design of sequential controlled clinical trials. Journal of the American Statistical Association, 73(363), 559–563.
Rosenberger, W.F. & Lachin, J.M. (2015). Randomization in Clinical Trials: Theory and Practice. Wiley.
Hu, F. & Rosenberger, W.F. (2006). The Theory of Response-Adaptive Randomization in Clinical Trials. Wiley.
Berry, S.M., Carlin, B.P., Lee, J.J. & Muller, P. (2010). Bayesian Adaptive Methods for Clinical Trials. CRC Press.
Pallmann, P., Bedding, A.W., Choodari-Oskooei, B., et al. (2018). Adaptive designs in clinical trials: why use them, and how to run and report them. BMC Medicine, 16, 29.