Why Constrain an Adaptive Allocation?
A response-adaptive rule may increasingly assign participants to the treatment that appears more effective as outcomes accrue. This can create a highly unequal allocation, particularly when early responses are noisy. An allocation cap limits how large an assignment probability can become; a minimum exploration probability ensures that each treatment retains some chance of assignment. These constraints can keep the less-favored arm from receiving very few participants, but they do not guarantee balanced realized sample sizes or higher power.
The effects depend on what is being capped, how probabilities evolve, and how the trial is analyzed. In particular, a bound on each participant’s assignment probability is not the same as a fixed bound on the final number assigned to an arm. This distinction matters for both interpretation and operating-characteristic calculations.
In a two-arm trial, if the probability of assignment to either arm is capped at \(c\), then both arm probabilities lie between \(1-c\) and \(c\), assuming \(c\geq 0.5\). A minimum probability \(\epsilon\) for each arm instead gives \(p_t\in[\epsilon,1-\epsilon]\). These descriptions coincide when \(\epsilon=1-c\). In a trial with more than two arms, an upper cap and a minimum probability are generally distinct constraints: limiting one arm’s probability does not specify how probability is distributed among all the others.
A realized-count constraint is different. For example, a design might use a rule that prevents either arm from exceeding a specified share of total enrollment. Such a rule controls cumulative counts directly, whereas probability bounds govern the randomization at each assignment. Even a constant assignment probability does not ensure an exact final share.
What the Constraints Change
If a response-adaptive rule would assign nearly everyone to the apparently better treatment, an upper cap limits that concentration. A floor also prevents an arm’s assignment probability from approaching zero. The less-favored arm can therefore continue to accrue participants, which can improve the information available about its outcomes and keep the trial from becoming nearly uninformative about that arm.
The trade-off is that the floor assigns more participants to an arm that the current rule regards as less promising. If that arm is truly inferior, the constraint increases exposure to it relative to the unconstrained rule. Conversely, the cap reduces assignments to the arm currently being favored. These are direct consequences of the assignment probabilities, not guarantees about the patients’ outcomes: an arm that looks inferior during the trial need not be inferior in truth.
The constraint can also change power. For a fixed total sample size and a common outcome variance, more balanced arm counts generally give a more precise difference in sample means than very unequal counts. But a cap need not produce equal counts, and any remaining imbalance can still reduce precision relative to equal allocation. Earlier material on power under response-adaptive randomization and on sensitivity to treatment-effect magnitude motivates examining both the allocation distribution and the effect size, rather than assuming one constraint will always help.
Worked Example: A Probability Floor and Expected Assignments
Suppose a hypothetical two-arm trial enrolls 40 participants. The first 10 are assigned with probability \(0.5\) to each arm. For the remaining 30, an adaptive rule would assign to \(A\) with probability \(0.05\) at every step. Assume, for this illustration, that these probabilities are fixed as stated. Compare the expected number assigned to \(A\) with and without a floor of \(0.15\) on its assignment probability. Suppose the rule regards \(A\) as the less promising treatment.
For an assignment indicator \(I_t\), equal to 1 if participant \(t\) is assigned to \(A\) and 0 otherwise, \(E(I_t)=P(A_t=A)\). Thus the expected total number assigned to \(A\) is the sum of its assignment probabilities. Without a floor,
With the floor, the later assignment probability becomes \(0.15\), so
The floor raises the expected number assigned to the less-promising arm from 6.5 to 9.5 participants. This increases exposure to that arm under the stated hypothetical rule and makes very low enrollment in \(A\) less likely. It does not force exactly 9.5 participants into \(A\): realized counts are integers and vary randomly. The calculation relies on the specified assignment probabilities; in a genuinely adaptive trial, probabilities usually depend on the observed history, so their expected sum must be evaluated under that evolving process.
Precision and the Integer Allocation Problem
A useful benchmark is the variance of the difference in sample means under a fixed allocation. Suppose outcomes are independent within and between arms, each arm has a normal outcome distribution with the same known variance \(\sigma^2\), and the allocation counts \(n_A\) and \(n_B\) are fixed in advance, with \(n_A+n_B=n\). For \(\hat{\Delta}=\bar{x}_A-\bar{x}_B\), the variance is
For fixed \(n\), this variance is smallest at equal allocation when \(n\) is even. When \(n\) is odd, equal allocation is impossible; the variance-minimizing integer allocation is the closest possible split, with arm counts differing by one. Here is a short justification. Since \(n_B=n-n_A\), the product \(n_A n_B\) is largest when the two integer counts are as close as possible. Also,
Because the numerator is fixed, maximizing the product minimizes the variance. This argument also shows why a cap that moves a very unequal allocation toward balance can improve precision relative to that allocation, without implying that the cap achieves the minimum possible variance.
The benchmark has important assumptions: outcomes are independent, the common variance model is appropriate, and the counts are fixed rather than selected in response to observed outcomes. In an adaptive design, counts are random and can depend on earlier outcomes. The fixed-count expression is informative for understanding the allocation mechanism, but simply substituting expected counts into it does not in general give the exact variance or power of an adaptive procedure.
Worked Example: The Best Integer Split for Five Participants
Consider a hypothetical fixed-allocation study with \(n=5\) participants and independent normal outcomes with common variance \(\sigma^2=1\) outcome-unit squared. Compare the closest-to-equal split, \(n_A=2,n_B=3\), with the more imbalanced split, \(n_A=1,n_B=4\).
For the \(2/3\) split, the variance of the difference in means is
For the \(1/4\) split, it is
Of the integer allocations, \(2/3\) and \(3/2\) are closest to equal and have the smaller variance. The continuous calculation would suggest the impossible allocation \(2.5/2.5\); it must not be treated as an available split. This example illustrates the integer qualification: equal allocation minimizes variance when the total is even, while the closest possible split does so when it is odd.
Consequences for Power
For a fixed allocation under the same normal, common-variance assumptions, suppose the null hypothesis is \(H_0:\mu_A-\mu_B=0\), the alternative is two-sided, and \(\sigma\) is known. The standardized statistic has standard error \(SE=\sigma\sqrt{1/n_A+1/n_B}\). Under a true difference \(\Delta=\mu_A-\mu_B\), its mean is \(\Delta/SE\). The two-sided \(z\)-test rejects when the absolute statistic exceeds \(z_{1-\alpha/2}\); its power is the probability of rejection under that true difference.
The following calculation is a fixed-allocation benchmark, not an exact power calculation for an outcome-adaptive rule. It shows how an allocation cap that limits extreme imbalance can affect precision and approximate power.
Worked Example: A Cap, the Standard Error, and Approximate Power
In a hypothetical trial, assume 100 participants, independent normal outcomes with known common standard deviation \(\sigma=10\) outcome units, a true difference \(\Delta=3\) outcome units, and a two-sided test at \(\alpha=0.05\). Compare fixed allocations of \(50/50\), \(20/80\), and \(5/95\). The latter is an extreme allocation; a cap that limits either arm to at most 80% of assignments would exclude it.
For \(50/50\), the standard error and standardized true difference are
For \(20/80\),
For \(5/95\),
For this test, power is \(P(|Z|>1.96)\) when \(Z\) is normal with mean \(\Delta/SE\) and variance 1. Thus the approximate powers for the three allocations are, respectively, \(0.323\), \(0.224\), and \(0.100\). For example, for the \(20/80\) allocation,
Under these assumptions, the \(20/80\) cap benchmark has greater power than the \(5/95\) allocation, but less than the balanced allocation. The result does not show that a cap always raises power: it depends on the comparison allocation, outcome model, effect size, and analysis. Nor is \(0.224\) the exact power of a response-adaptive design whose random counts depend on outcomes.
Diagnostics and Comparisons
Assess a constraint by examining its operating characteristics, not only its nominal bound. Useful summaries include the distribution of final allocation counts, the frequency of very small arm sizes, the range and trajectory of assignment probabilities, and power under the planned analysis. For an adaptive rule, these quantities can be estimated by repeatedly simulating the full design under specified treatment effects and outcome assumptions. A single expected allocation can conceal substantial variation across trials.
Check that the implementation applies the constraint at every intended randomization. If the floor is meant to be \(\epsilon\), verify that neither arm’s probability falls below it after each update. Also inspect whether the probability is being capped or the realized cumulative allocation is being controlled; those are different design features. If outcomes are delayed or probabilities are updated on a different schedule, the relevant histories and update times must be reflected in the simulation.
| Approach | What it controls | Main implication |
|---|---|---|
| Unconstrained adaptive probabilities | No stated floor or cap on each assignment probability | Can concentrate allocation strongly, especially after noisy early evidence. |
| Probability cap or minimum exploration | Allowed assignment probabilities at each step | Preserves some probability of assigning to each arm but does not fix final counts. |
| Fixed equal randomization | Uses the same probability, usually one-half, at each assignment | Provides a balanced allocation in expectation, but realized counts can still differ. |
| Realized-count constraint | Cumulative allocation shares or counts | Directly restricts final or interim imbalance, but is not equivalent to a per-assignment probability bound. |
A frequent mistake is to reason, “The probability floor is 0.2, so at least 20% of participants will be assigned to that arm.” The floor means each eligible participant has at least a 0.2 chance of assignment to that arm at that randomization; chance variation means the realized proportion can be below 20%. The correction is to distinguish a probability bound from a deterministic count guarantee and to report the distribution of final counts.
Another mistake is to conclude that a cap must increase power because it reduces imbalance. It may improve precision relative to an especially extreme allocation, but if the cap still permits substantial imbalance, power can remain below that of balanced allocation under the fixed-count, equal-variance benchmark. Adaptive power also depends on random allocation counts and the validity and performance of the analysis procedure. A cap alone does not establish type I error control or solve problems caused by an unsuitable analysis.
Check Your Understanding
Answer each question, then compare with the solutions.
- In a two-arm trial, every participant has probability between \(0.1\) and \(0.9\) of assignment to \(A\). Does this guarantee that at least 10% of the realized sample will receive \(A\)?
- For \(n=7\) participants and equal common outcome variances, which integer split minimizes the variance of the difference in sample means?
- With 60 participants, compare the expected number assigned to \(A\) if its assignment probability is fixed at \(0.1\) versus fixed at \(0.2\). What does a floor of \(0.2\) change?
- Under the fixed-count, equal-variance benchmark, why can a cap improve precision compared with a \(5/95\) split while still having less power than a \(50/50\) split?
Solutions. (1) No. The bound applies at each randomization, while realized counts vary by chance; it is not a deterministic final-share constraint. (2) The closest-to-equal split, \(3/4\) or \(4/3\). The product of the arm counts is largest for this split, minimizing \(n/(n_A n_B)\). (3) The expectations are \(60(0.1)=6\) and \(60(0.2)=12\) assignments to \(A\). The floor raises the expected assignments in this fixed-probability example, but does not force exactly 12. (4) A \(5/95\) split has a larger value of \(1/n_A+1/n_B\), and therefore a larger standard error than a less imbalanced allocation. But the \(50/50\) split minimizes that quantity for an even total sample size, so a capped allocation that remains imbalanced can still have lower power than equal allocation under the stated benchmark.