Ethical Trade-Offs in Adaptive Allocation

Why Adaptive Allocation Creates an Ethical Trade-Off

In a randomized clinical trial, an allocation rule determines how likely each eligible participant is to receive each available arm. When response information accumulates during the trial, those probabilities may change. If one treatment appears more promising, assigning more participants to it may improve expected outcomes for participants enrolled later. But concentrating assignments can reduce what the trial learns about the other arm, weaken treatment comparisons, and make estimates less precise. The ethical question is therefore not simply whether adaptation is possible, but how to balance expected benefit now against the value of learning.

These goals can point in different directions. A rule aimed only at the expected outcome of the next participant might assign everyone to the arm with the highest current posterior mean. A rule aimed at comparing treatments may preserve allocation to both arms, even if one currently looks better. Neither goal automatically overrides the other. Their relative importance is a design choice that should be justified before the trial, not selected opportunistically after seeing results.

Definition: Let \(q_{1,t}\) be the probability that the next eligible participant is assigned to treatment arm 1, conditional on the information available before that assignment. With two active arms, the probability of assignment to arm 0 is \(q_{0,t}=1-q_{1,t}\). An ethical allocation objective specifies how the design weighs expected outcomes for participants enrolled during the trial against information useful for estimating treatment effects or making later decisions.

An assignment probability is not a probability that an arm is effective, nor is it itself a measure of evidence. For example, \(q_{1,t}=0.70\) means that the randomization mechanism assigns the next participant to arm 1 with probability 0.70, given the current trial history. It does not mean there is a 70% probability that arm 1 is superior. A rule may use evidence about superiority to determine assignment probabilities, but those are distinct quantities.

State the Objective Before Choosing the Rule

For a simple illustration, suppose the outcome is binary and a response is considered beneficial. Let \(p_1\) and \(p_0\) denote the response probabilities on arms 1 and 0. Given the trial data available at time \(t\), their posterior means are the current model-based expectations of those response probabilities. Under the working model used for the allocation decision, the expected response probability for the next participant is

q_{1,t}E(p_1\mid \mathcal{D}_t) +(1-q_{1,t})E(p_0\mid \mathcal{D}_t),

where \(\mathcal{D}_t\) denotes the information available before the next assignment. If the only objective were to maximize this expression, the rule would assign the next participant to the arm with the larger posterior mean, with no allocation to the other arm unless the means were tied. This is sometimes called a greedy objective. It favors the immediate expected outcome under the model, but it can sacrifice learning and may respond strongly to uncertain early evidence.

A trial can instead preserve allocation for learning by imposing minimum assignment probabilities, by using a more balanced rule, or by explicitly combining a participant-benefit target with a learning target. Such choices do not remove the trade-off; they make it visible and governable. The design should specify how probabilities are calculated, how often they may change, and what safeguards limit extreme allocation. It should also state whether the modeled response is a suitable proxy for participant benefit. A response endpoint alone may not represent all clinically relevant outcomes, and a favorable posterior mean is not a guarantee of benefit for an individual.

Worked Example: Combining a Benefit Target With a Learning Target

Hypothetical trial. Suppose each arm has a uniform Beta(1,1) prior response probability. At an interim look, arm 1 has one response among one observed participant, and arm 0 has no participants with observed outcomes. Under independent Bernoulli outcomes within arms, the posterior distributions are \(p_1\mid\mathcal{D}_t\sim\mathrm{Beta}(2,1)\) and \(p_0\mid\mathcal{D}_t\sim\mathrm{Beta}(1,1)\). The posterior means are \(2/3\) and \(1/2\), respectively. The small amount of information is important context: these estimates are uncertain.

For this illustration, define a benefit target as the posterior probability that arm 1 has the larger response probability. With the stated posterior distributions,

P(p_1>p_0\mid\mathcal{D}_t) =\int_0^1 2x\left(\int_0^x 1\,dy\right)dx =\int_0^1 2x^2\,dx =\frac{2}{3}.

Consider a transparent design choice that blends this target with equal allocation, \(q_{\mathrm{learn}}=0.50\). Let \(\lambda=0.60\) represent the prespecified weight given to the benefit target, and use

q_{1,t}=\lambda P(p_1>p_0\mid\mathcal{D}_t) +(1-\lambda)q_{\mathrm{learn}}.

Substituting the values gives \(q_{1,t}=0.60(2/3)+0.40(0.50)=0.60\). Thus the next participant has a 60% assignment probability for arm 1 and a 40% probability for arm 0. Under the posterior means, the expected response probability for that participant is \(0.60(2/3)+0.40(1/2)=0.60\). Equal allocation would give \(0.50(2/3)+0.50(1/2)=0.5833\), while the greedy choice of arm 1 alone would give \(0.6667\).

This calculation illustrates a policy, not a universally optimal ethical formula. The blend gives some priority to the arm more likely to be superior while retaining allocation to arm 0. The value \(\lambda=0.60\) is a design judgment; it is not estimated from the data or dictated by Bayesian analysis. With only one observed outcome, the posterior evidence is particularly fragile, which is a reason to scrutinize any policy that would rapidly concentrate assignments.

Learning Has a Cost and a Value

The learning objective is not simply “equal numbers.” For a conventional two-arm comparison of binary response proportions, a useful approximation shows how allocation affects the variance of the estimated difference. Suppose there are \(N\) participants with outcomes observed, \(n_1\) assigned to arm 1 and \(n_0=N-n_1\) assigned to arm 0. Under independent Bernoulli outcomes with response probabilities \(p_1\) and \(p_0\), the variance of the difference in sample response proportions is

\operatorname{Var}(\hat p_1-\hat p_0) =\frac{v_1}{n_1}+\frac{v_0}{n_0}, \qquad v_1=p_1(1-p_1),\quad v_0=p_0(1-p_0).

For this fixed-sample comparison, treat the variances as known and let \(q=n_1/N\), so \(n_0=(1-q)N\). Then the variance as a function of the allocation proportion is \(N^{-1}\{v_1/q+v_0/(1-q)\}\). Differentiating with respect to \(q\), setting the derivative to zero, and solving gives

-\frac{v_1}{q^2}+\frac{v_0}{(1-q)^2}=0 \quad\Longrightarrow\quad \frac{\sqrt{v_1}}{q}=\frac{\sqrt{v_0}}{1-q} \quad\Longrightarrow\quad q=\frac{\sqrt{v_1}}{\sqrt{v_1}+\sqrt{v_0}}.

This stationary point minimizes the variance: the second derivative of \(v_1/q+v_0/(1-q)\) is \(2v_1/q^3+2v_0/(1-q)^3\), which is positive for \(0<q<1\) when both variances are positive. Equal allocation minimizes this variance exactly when \(v_1=v_0\). When the variances are close, equal allocation may be approximately optimal; it is not generally the exact variance-minimizing choice. This calculation concerns precision for the stated comparison, not the full ethical value of the trial or the optimality of an adaptive rule.

Worked Example: Comparing Balanced and Variance-Minimizing Allocation

Hypothetical planning values. Suppose the binary response probabilities are \(p_1=0.60\) and \(p_0=0.50\), and the total sample size is \(N=200\). Assume independent outcomes and use these probabilities as fixed planning values. The arm-specific outcome variances are \(v_1=0.60(0.40)=0.24\) and \(v_0=0.50(0.50)=0.25\). The variance-minimizing proportion assigned to arm 1 is

q=\frac{\sqrt{0.24}}{\sqrt{0.24}+\sqrt{0.25}} =\frac{0.4899}{0.4899+0.5000} \approx 0.4949.

This corresponds to approximately \(98.98\) of 200 participants on arm 1 and \(101.02\) on arm 0 in the continuous calculation; actual assignments must be integer-valued. At equal allocation, the variance is \(0.24/100+0.25/100=0.004900\). At the continuous optimum, it is approximately \(0.24/98.98+0.25/101.02=0.004900\) to the shown precision. The difference is very small here because the two arm variances are close, but the exact optimum is not \(q=0.50\). This calculation concerns variance under the specified fixed-allocation model; it does not establish an ethical allocation policy for an adaptive trial.

Allocation Bounds as an Explicit Ethical Safeguard

A simpler policy can begin with a direct participant-benefit objective and then limit how concentrated assignments may become. Suppose the posterior means at an interim look are \(E(p_1\mid\mathcal{D}_t)=2/3\) and \(E(p_0\mid\mathcal{D}_t)=1/2\). If the design requires the allocation probability to each arm to be at least 0.25, then \(q_{1,t}\) must lie between 0.25 and 0.75. Within these limits, assigning more participants to arm 1 increases the model-based expected response for the next participant, so the rule selects the upper bound. The bound does not claim that arm 1 is certainly better; it protects continued assignment to arm 0 while allowing some preference for arm 1.

Worked Example: Applying a Prespecified Allocation Bound

Hypothetical trial. Use the posterior means above and suppose 40 participants remain to be assigned. A prespecified rule restricts each arm’s assignment probability to the range 0.25 to 0.75, and within that range favors the arm with the larger posterior mean. Since \(2/3>1/2\), the selected probability is \(q_{1,t}=0.75\), with \(q_{0,t}=0.25\). This means a 75% chance of assignment to arm 1 for each assignment made under this rule; it does not mean a 75% posterior probability that arm 1 is better.

The expected number assigned to arm 1 among 40 assignments is \(40(0.75)=30\), and the expected number assigned to arm 0 is \(40(0.25)=10\). Under the current posterior means, the expected number of responses is \(30(2/3)+10(1/2)=20+5=25\). For comparison, equal allocation gives an expected \(20(2/3)+20(1/2)=13.33+10=23.33\) responses, while assigning all 40 to arm 1 gives \(40(2/3)=26.67\). These are expectations under the stated posterior model, not guaranteed outcomes. The bound trades some modeled immediate response expectation for continued allocation to arm 0 and the possibility of learning about it.

How to Judge Whether the Trade-Off Is Defensible

A defensible allocation policy needs more than a mathematical rule. The rule should be prespecified and understandable to the trial team, and its probability changes should be reviewable. The design rationale should identify what counts as participant benefit, what learning is needed, and why the chosen minimum or maximum assignment probabilities are appropriate. In a real clinical setting, the protocol also needs to address relevant uncertainties and clinical obligations that a simplified response model cannot represent. Statistical efficiency, expected responses, and ethical acceptability are related considerations, but they are not interchangeable conclusions.

Several warning signs merit attention when evaluating a proposed policy:

The objective is implicit. If the design cannot explain whether it prioritizes immediate outcomes, precise comparisons, or a specified balance, it is difficult to assess whether its probabilities are appropriate.
A probability is described as proof. A high assignment probability may result from a policy choice and does not establish a high posterior probability of superiority.
Early evidence is treated as stable. When few outcomes inform the posteriors, apparent differences may be uncertain. A policy that shifts sharply on limited information can direct many assignments according to chance variation.
“Balanced is ethical” is used without qualification. Balance can support learning and comparison, but it does not by itself ensure that a trial is ethically acceptable. Conversely, an unbalanced allocation is not automatically unethical; its justification depends on the objective, evidence, safeguards, and context.
A proxy outcome stands in for all benefit. A response probability may omit other outcomes relevant to participants. An allocation objective based on that endpoint should be presented as a limited model-based objective, not a complete ethical judgment.

The variance calculation above also has limits. It assumes independent binary outcomes and a fixed total sample size, and it treats the response variances as known. In an adaptive trial, the probabilities depend on accumulating data; realized allocation counts and estimator properties can therefore differ from the fixed-allocation calculation. The calculation is still useful for explaining why allocation affects information, but it is not a substitute for assessing the specific adaptive design.

Key takeaway: Ethical allocation is a deliberate balance between expected outcomes for participants assigned now and the value of learning from the trial. State that objective explicitly, calculate the resulting probabilities transparently, and do not confuse an allocation probability with evidence of treatment superiority.

Check Your Understanding

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

If \(q_{1,t}=0.70\), what does this number mean, and what does it not mean?
For posterior means \(E(p_1\mid\mathcal{D}_t)=0.55\) and \(E(p_0\mid\mathcal{D}_t)=0.50\), calculate the expected response probability under equal allocation and under \(q_{1,t}=0.60\).
Under the fixed-sample variance calculation, when does equal allocation minimize the variance of the difference in two sample response proportions?
A rule assigns all participants to the arm with the larger posterior mean. Name one objective it favors and one potential cost.

Solutions. (1) It is the conditional randomization probability of assigning the next eligible participant to arm 1 under the rule. It is not the probability that arm 1 is superior. (2) Under equal allocation, the expectation is \(0.50(0.55)+0.50(0.50)=0.525\). With \(q_{1,t}=0.60\), it is \(0.60(0.55)+0.40(0.50)=0.530\). These are model-based expected response probabilities for the next assignment. (3) Equal allocation minimizes the stated variance exactly when the two arm outcome variances are equal, \(p_1(1-p_1)=p_0(1-p_0)\). (4) It favors the model-based expected outcome for participants assigned immediately. A potential cost is reduced allocation to the other arm, which can limit learning and weaken the treatment comparison.