Response-Adaptive Randomization in a Two-Arm Trial

Why Adapt Randomization?

In a two-arm trial, response-adaptive randomization uses accumulating outcome data to change the probabilities of assigning future participants to treatment and control. If early evidence suggests that treatment has the higher response rate, a design may assign more participants to treatment. This is a change to the assignment mechanism: it does not change the meaning of a response, make a treatment effective, or guarantee that more participants will benefit.

A practical rule must also specify how much allocation can change. Assigning nearly everyone to the arm that currently looks better can leave too few control participants, and early evidence may be noisy. The example rule below uses the posterior probability that treatment is better, but caps its influence so that neither arm receives less than a prespecified fraction of assignments.

Model and Allocation Rule

Definition: Let \(\theta_T\) and \(\theta_C\) be the probabilities of a favorable binary response under treatment and control. Conditional on these probabilities, assume responses are independent Bernoulli observations within each arm. Use independent priors \(\theta_T\sim\operatorname{Beta}(a_T,b_T)\) and \(\theta_C\sim\operatorname{Beta}(a_C,b_C)\), where each first parameter describes prior successes and each second parameter prior failures. At an interim look, let \(y_j\) be the number of responses and \(n_j\) the number of participants with observed outcomes in arm \(j\), for \(j\in\{T,C\}\). The posterior is \(\theta_j\mid\text{data}\sim\operatorname{Beta}(a_j+y_j,b_j+n_j-y_j)\).

This model assumes the response definition is consistent across arms, responses are observed before the next allocation decision, and the chosen priors are suitable for the clinical context. The independent Bernoulli model also treats participants’ outcomes as conditionally independent given their arm response probability. The rule is written for an outcome where a larger response probability is better; an endpoint where smaller values are preferable requires defining the favorable event appropriately.

Define the posterior superiority probability \(q=P(\theta_T>\theta_C\mid\text{observed data})\). For a chosen allocation floor \(\varepsilon\), with \(0<\varepsilon<0.5\), this tutorial uses:

p_T=\min\{1-\varepsilon,\max\{\varepsilon,q\}\}, \qquad p_C=1-p_T.

At the next assignment, treatment is selected with probability \(p_T\) and control with probability \(p_C\). Thus each arm has allocation probability at least \(\varepsilon\). The posterior probability \(q\) is not itself an allocation probability in every circumstance: when it falls below the floor or above the cap, the rule replaces it with the boundary value. In a continuous beta model, the probability of an exact tie between the two response probabilities is zero.

Updating and Calculating Superiority

The posterior update follows directly from the likelihood. For arm \(j\), if its prior density is proportional to \(\theta_j^{a_j-1}(1-\theta_j)^{b_j-1}\), then \(y_j\) responses and \(n_j-y_j\) nonresponses contribute likelihood proportional to \(\theta_j^{y_j}(1-\theta_j)^{n_j-y_j}\). Multiplying these gives:

\theta_j\mid\text{data} \ \propto\ \theta_j^{a_j+y_j-1}(1-\theta_j)^{b_j+n_j-y_j-1},

which is the beta density with parameters \(a_j+y_j\) and \(b_j+n_j-y_j\). To obtain \(q\), one can integrate the treatment posterior density against the control posterior cumulative distribution function:

q=\int_0^1 f_T(t\mid\text{data})F_C(t\mid\text{data})\,dt.

This expression sums, over possible treatment response probabilities \(t\), the posterior probability density at \(t\) times the probability that the control response probability is below \(t\). The calculation may be done by numerical integration or an appropriate exact calculation. The following examples show the integral when the beta shapes make it simple.

Worked Example: A Low Superiority Probability and the Control Floor

Suppose both arms begin with independent \(\operatorname{Beta}(1,1)\) priors. At an interim look, one treatment participant has an observed nonresponse, while one control participant has an observed response. The posterior distributions are:

\theta_T\mid\text{data}\sim\operatorname{Beta}(1,2), \qquad \theta_C\mid\text{data}\sim\operatorname{Beta}(2,1).

For the treatment posterior, the density is \(2(1-t)\). For the control posterior, the cumulative distribution function is \(t^2\). Therefore:

q=\int_0^1 2(1-t)t^2\,dt =2\left(\frac{1}{3}-\frac{1}{4}\right) =\frac{1}{6}\approx 0.1667.

With \(\varepsilon=0.20\), the unbounded value \(q\) is below the allocation floor. The actual probabilities are \(p_T=0.20\) and \(p_C=0.80\). The control arm is therefore retained at no less than 20% of assignments, while treatment still has a nonzero chance of assignment.

As a check on the direction of the calculation, reverse the posteriors: let treatment be \(\operatorname{Beta}(2,1)\) and control be \(\operatorname{Beta}(1,2)\). These continuous distributions are mirror images, so the two strict superiority probabilities sum to 1. The reversed case has \(q=1-\frac{1}{6}=\frac{5}{6}\), not \(0.5\). With the same floor, its allocation probabilities are \(p_T=0.80\) and \(p_C=0.20\).

Worked Example: A Posterior Calculation From Unequal Response Counts

Suppose independent \(\operatorname{Beta}(1,1)\) priors are used for both arms. At one look, 3 of 4 treatment participants and 1 of 2 control participants have responded. The posteriors are \(\theta_T\mid\text{data}\sim\operatorname{Beta}(4,2)\) and \(\theta_C\mid\text{data}\sim\operatorname{Beta}(2,2)\).

The treatment density is \(20t^3(1-t)\), and the control cumulative distribution function is \(3t^2-2t^3\). Substitution into the integral gives:

\begin{aligned} q &=\int_0^1 20t^3(1-t)(3t^2-2t^3)\,dt\\ &=20\int_0^1(3t^5-5t^6+2t^7)\,dt\\ &=20\left(\frac{3}{6}-\frac{5}{7}+\frac{2}{8}\right) =\frac{5}{7}\approx 0.7143. \end{aligned}

For \(\varepsilon=0.20\), this value is between the floor and cap, so \(p_T=5/7\approx0.7143\) and \(p_C=2/7\approx0.2857\). The conclusion is about the next assignment probabilities under this rule, not a declaration that treatment is superior or a guarantee that a treatment-assigned participant will respond.

Following the Rule Across Interim Looks

A response-adaptive design updates the posterior only when eligible outcomes have been observed, then uses the updated allocation probabilities for subsequent assignments. In practice, the protocol must state when looks occur, which outcomes are available at each look, and how assignments are generated. The simple sequence below shows how the probabilities can change as new outcomes arrive.

Worked Example: Updating Allocation After a New Cohort

Assume independent \(\operatorname{Beta}(1,1)\) priors and an allocation floor of \(\varepsilon=0.20\). At the first look, treatment has 0 responses among 1 participant and control has 1 response among 1 participant. The posteriors are \(\operatorname{Beta}(1,2)\) for treatment and \(\operatorname{Beta}(2,1)\) for control. As calculated above, \(q=1/6\), so the next assignments use probabilities \(p_T=0.20\) and \(p_C=0.80\).

Suppose that a subsequent cohort is assigned under those probabilities and its outcomes are observed before the next look. The cohort contains one treatment participant with a response and one control participant with a nonresponse. Across all observed participants, treatment now has 1 response among 2, and control has 1 response among 2. Updating from the original priors gives:

\theta_T\mid\text{data}\sim\operatorname{Beta}(2,2), \qquad \theta_C\mid\text{data}\sim\operatorname{Beta}(2,2).

The posterior distributions are identical and continuous, so each arm has probability \(1/2\) of having the larger response probability. Thus \(q=0.5000\), and the floor does not bind: \(p_T=p_C=0.5000\). The probability of the particular cohort assignments and outcomes in this illustration is not what determines the allocation rule; the rule uses the observed data available before the next assignment.

Implementation, Diagnostics, and Alternatives

A clear operational sequence helps prevent accidental use of information that was not available when a participant was randomized.

Specify the model and rule.
Record the response definition, priors, interim timing, allocation floor, and exact calculation of \(q\) in advance.

Update from observed outcomes.
At each planned look, count responses and nonresponses by arm and calculate the two posterior distributions.

Calculate and constrain allocation.
Compute \(q\), apply the floor and cap, and verify that \(p_T+p_C=1\) and both probabilities are at least \(\varepsilon\).

Randomize and record.
Use the stated probabilities for future assignments and retain the assignment probability and decision-time data for each randomization.

Useful checks include confirming that successes plus failures equal the number of observed outcomes in each arm, that posterior parameters increase by the correct counts, and that the calculated allocation probabilities obey the floor. Reversing the arm labels should reverse the superiority probability: if there are no ties, \(P(\theta_T>\theta_C\mid\text{data})+P(\theta_C>\theta_T\mid\text{data})=1\). A failure of this check can reveal a reversed density or cumulative distribution function, an incorrect beta parameter, or a numerical integration error.

The main alternative is fixed randomization, such as assigning each participant with probability \(1/2\) to each arm, regardless of interim outcomes. Fixed allocation is simpler to operate and does not change assignment ratios in response to early random variation. Response-adaptive allocation can increase the share assigned to the arm with greater posterior support, but the probabilities can be sensitive to the prior and to a small number of early outcomes. Neither approach is uniformly preferable; the choice depends on the trial’s objectives, feasibility, and analysis plan.

An allocation floor preserves a minimum control allocation, but it does not ensure a large or precise control sample, nor does it eliminate ethical or inferential concerns. With delayed, missing, or selectively observed outcomes, a look may not represent all participants enrolled so far. Strongly changing allocation probabilities can also complicate trial operations and interpretation. The rule should therefore be evaluated as part of the full design, and the final analysis must account for the trial’s data and assignment process as specified in its analysis plan.

Common Pitfalls and Limits

Confusing \(q\) with certainty. A posterior superiority probability is conditional on the model, prior, and observed data. It is not proof of a treatment effect.
Using \(q\) without the allocation constraint. If \(q=0.03\) and \(\varepsilon=0.20\), the rule assigns to treatment with probability 0.20, not 0.03. Control then receives probability 0.80.
Updating before outcomes are available. A participant’s outcome cannot be used to determine that same participant’s assignment. Only information available before a randomization may govern its probability.
Assuming adaptation removes uncertainty. Early response counts can be unstable, and the posterior depends on the prior and model. Adaptation changes who is assigned next; it does not make the evidence more certain by definition.

For illustration, the beta-binomial rule is deliberately simple. It may be unsuitable if response probabilities vary materially over time or across important patient groups, if outcomes are not observed promptly, or if the clinical objective requires balancing more than response rate. Those features require a design and analysis model that represents them rather than relying on the simple two-rate calculation.

Check Your Understanding

Answer each question, then compare your reasoning with the solutions.

With a \(\operatorname{Beta}(1,1)\) prior and 7 responses among 10 treatment participants, what is the treatment posterior?
If the posteriors are treatment \(\operatorname{Beta}(2,1)\) and control \(\operatorname{Beta}(1,2)\), what is \(q\), and what are the allocation probabilities when \(\varepsilon=0.15\)?
If \(q=0.08\) and \(\varepsilon=0.20\), what are \(p_T\) and \(p_C\)?
Why must the allocation probability for a participant be calculated before that participant’s response is observed?

Solutions. (1) The posterior is \(\operatorname{Beta}(1+7,1+10-7)=\operatorname{Beta}(8,4)\). (2) The mirror-image posteriors give \(q=5/6\approx0.8333\). This is below the cap \(0.85\), so \(p_T=5/6\) and \(p_C=1/6\approx0.1667\); both exceed the floor. (3) Since \(0.08<0.20\), the floor gives \(p_T=0.20\) and \(p_C=0.80\). (4) Randomization must use only information available at the time of assignment. Using a future response would make the stated assignment mechanism impossible to implement and would improperly condition assignment on an outcome not yet observed.