When Adaptation Leaves an Active Arm Behind
Response-adaptive allocation changes treatment assignment probabilities as trial data accumulate. That flexibility can become a problem when the rule assigns an active arm a probability close to zero. The arm has not been formally dropped, yet it may receive few or no further participants. The result can be weak comparative information: estimates involving that arm may be imprecise, and the data may provide little practical support for a comparison.
The central diagnostic question is not simply whether allocation is unequal. Unequal allocation may be deliberate and useful. The question is whether the rule leaves each arm that is still intended for comparison with a meaningful chance of assignment, across the histories and time periods that matter. There is no universal cutoff that defines “overly deterministic”; the threshold depends on the design, expected enrollment, outcome, and inferential goals.
The conditioning matters. A design may assign an arm fairly often on average but give it almost no chance after particular histories. Conversely, probabilities may vary from one update to the next without ever becoming too concentrated. Assess the probability path, not just a single overall average. Also distinguish a low probability for an arm that remains active from probability zero after a prespecified rule has formally dropped that arm; those are different design states.
Summarizing How Concentrated an Allocation Is
For a decision with \(K\) active arms, the entropy of the allocation probabilities is \(H_t=-\sum_{a\in\mathcal{A}_t}p_{a,t}\log(p_{a,t})\), using the convention \(0\log(0)=0\). Entropy is measured in nats when the natural logarithm is used. It is largest, \(\log(K)\), when all active arms have equal probability, and it is zero when assignment is deterministic. Because its maximum depends on the number of active arms, normalized entropy \(H_t/\log(K)\) is useful when \(K>1\): it ranges from 0 to 1.
Entropy is a summary, not a pass/fail test. It can flag concentrated allocation, but it does not say which arm is receiving little support or whether the number of expected assignments is adequate. Report arm-specific probabilities alongside it. When the number of active arms changes, interpret entropy with that changing number in view.
Worked Example: Quantifying a Concentrated Three-Arm Decision
Suppose three arms remain active at a particular update, with hypothetical allocation probabilities 0.94, 0.05, and 0.01. The probabilities sum to 1, so they define a valid randomization decision. The smallest probability is only 1% at this decision.
The entropy is \(H=-[0.94\log(0.94)+0.05\log(0.05)+0.01\log(0.01)]\).
With three active arms, the maximum is \(\log(3)\approx1.0986\) nats. Thus the normalized entropy is \(0.2540/1.0986\approx0.2312\). This indicates a highly concentrated decision relative to equal allocation. The summary alone does not establish that the rule is unacceptable; the next step is to examine how often this probability occurs and how many assignments the arm can realistically expect.
From Small Probabilities to Weak Support
Expected enrollment and the probability of receiving no assignments give direct, horizon-specific diagnostics. If an arm has conditional allocation probability \(p\) at each of the next \(m\) assignments, then its expected number of assignments is \(mp\). Under those constant probabilities, the probability it receives none is \((1-p)^m\). If probabilities depend on the random history, the unconditional expected number is the sum of the expectations of the conditional probabilities. For a fixed, specified sequence of probabilities, the probability of no assignment is the product of the probabilities of not assigning that arm at each step.
The constant-probability formula makes the practical point clear: a probability can be positive at every assignment and still yield a substantial chance of no enrollment over a short horizon. Positivity in a mathematical sense is not the same as useful support in a finite trial. For adaptive rules, probabilities depend on observed history, so a single product calculated from an average probability is generally not an exact unconditional probability. Use the rule’s possible histories, or simulation under that rule, when a full design-level probability is required.
Worked Example: Chance of No Further Enrollment
Suppose an active arm has probability 0.02 of assignment at each of the next 30 decisions, with this probability held constant for this hypothetical calculation. Its expected number of assignments is \(30(0.02)=0.6\) participants. The probability of no assignments is
So even though the arm has a positive chance at every decision, there is about a 54.55% chance it receives no additional participant in this horizon. For comparison, if its probability were 0.20 at each decision, its expected assignments would be \(30(0.20)=6\), and the chance of none would be \(0.80^{30}\approx0.001238\). These calculations illustrate why “not dropped” is not enough to establish adequate future support.
In a real adaptive design, the probability may change after each outcome. The calculation above is valid only under the stated constant-probability assumption; for changing probabilities, the design should be assessed over its possible allocation paths.
Why Sparse Allocation Reduces Comparative Information
A simple continuous-outcome model shows the connection between arm counts and precision. Assume independent participant outcomes, with \(Y_i\sim N(\mu_a,\sigma^2)\) for a participant assigned to arm \(a\), a common known variance \(\sigma^2\), and no interference between participants. For fixed (or non-informatively selected) arm counts \(n_a\) and \(n_b\), the sample means estimate the arm-specific means, and the variance of their difference is \(\operatorname{Var}(\bar{Y}_a-\bar{Y}_b)=\sigma^2(1/n_a+1/n_b)\). This formula assumes both counts are positive. With outcome-adaptive assignment, conditioning on final counts can select on earlier outcomes, so the conditional sampling distribution and variance need not have this form; the analysis must account for the allocation mechanism. If an arm has no observations, its sample mean and this direct comparison are unavailable.
For fixed (or non-informatively selected) counts, independent sample means have variances \(\sigma^2/n_a\) and \(\sigma^2/n_b\); the variance of their difference is the sum of those variances. With outcome-adaptive assignment, conditioning on final counts can alter the outcome distribution, so this count-based formula need not apply. This is a transparent illustration of information loss, not a complete analysis of an adaptive trial. Covariate adjustment, time trends, unequal variances, missing outcomes, and the dependence induced by an adaptive rule can change the appropriate analysis. Still, a very small arm count cannot be repaired merely by using a more sophisticated estimator.
Worked Example: Comparing Precision Under Unequal Allocation
Consider a hypothetical two-arm trial with 100 participants, continuous outcomes, and common known standard deviation \(\sigma=10\) outcome units. Suppose the fixed counts are 5 in arm A and 95 in arm B. Under the stated model, the standard error of the difference in sample means for these fixed counts is
With equal observed counts of 50 and 50, the corresponding standard error would be \(10\sqrt{1/50+1/50}=10\sqrt{0.04}=2.000\) outcome units. In this illustration, the unequal allocation gives a standard error about \(4.588/2.000=2.294\) times as large for the arm-mean difference.
These calculations apply to the stated fixed counts and model; conditioning on counts produced by outcome-adaptive assignment may change the outcome distribution. Expected counts under an adaptive rule are not a substitute for observed counts, and plugging expected counts into this formula does not generally give the exact expected standard error. The example isolates the cost of having only five observations in one arm.
Practical Diagnostics and Safeguards
A useful assessment combines several views of the allocation rule. Inspect probabilities at each update, cumulative enrollment by arm, and the remaining enrollment horizon. Where possible, evaluate how often simulated trial paths produce very low probabilities, few assignments, or no assignments to an active arm. A mean allocation probability can hide a dangerous tail of paths; a single allocation-path plot can also hide the frequency of such paths across repeated trials.
- Arm-specific probability paths: Identify periods or histories in which an active arm’s probability approaches zero. Summaries should include minima or lower quantiles as well as averages.
- Expected and realized counts: Compare projected enrollment from the remaining probabilities with the amount needed for the planned comparisons. Track actual counts, since projected support may not materialize.
- Zero-assignment risk: Estimate the chance of no further assignments, or of ending with a count below a prespecified design-relevant amount, over a stated horizon.
- Concentration summaries: Use entropy or normalized entropy as a compact signal, while retaining arm-level values to show where the concentration lies.
- History sensitivity: Check whether low probabilities arise only after particular outcome histories or recur across many plausible histories.
A common safeguard is a probability floor: require \(p_{a,t}\geq\varepsilon\) for every arm that is still active and eligible, for a prespecified \(\varepsilon>0\). This prevents the rule from assigning an arbitrarily small chance to an active arm. The floor must be incorporated into the allocation rule itself; reporting a floor after the fact does not change the randomization mechanism. A floor also does not guarantee adequate enrollment: if \(\varepsilon\) is small or little enrollment remains, expected counts may still be inadequate.
A floor should be evaluated alongside the rule’s goals and operating characteristics. It can reduce concentration, but it also constrains how strongly allocation can respond to accumulated data. Other options include tempering the adaptation or retaining a fixed component in the assignment probabilities. Any such change should be evaluated under plausible outcome patterns, rather than assumed to solve every support problem.
How the Main Approaches Differ
| Approach | Allocation feature | Support implication |
|---|---|---|
| Fixed allocation | Probabilities are set in advance and do not respond to accumulating outcomes. | Known probabilities make projected support straightforward, though small fixed probabilities can still leave sparse arms. |
| Unconstrained response adaptation | Probabilities may become very unequal as the rule updates. | Can produce near-zero chances and history-dependent weak support for an arm that remains active. |
| Adaptive allocation with a probability floor | Probabilities respond to data but are constrained to stay above a chosen minimum for active arms. | Limits extreme concentration but does not itself ensure a sufficient final count or precise comparison. |
| Formal arm removal | An arm is no longer eligible under the design’s stopping or dropping rule. | Zero probability follows from a change in trial status, not merely from a near-zero adaptive probability. Comparisons after removal require careful attention to the data and estimand. |
These approaches are not interchangeable. Fixed allocation avoids outcome-driven changes in probabilities but is not automatically balanced in realized counts. An adaptive rule with a floor retains some response to accumulating information while protecting a minimum chance of assignment. Formal removal is a distinct design decision, not a convenient label for an arm that the randomization algorithm has nearly abandoned.
Common Mistakes and Limitations
- “The probability is above zero, so support is adequate.” A 1% chance repeated over a short horizon can still produce no assignments more often than not. Assess expected and realized counts, as well as the probability of zero or very sparse enrollment.
- “The average allocation is reasonable, so every path is reasonable.” Adaptive probabilities depend on history. Report their distribution across times and plausible paths, not only an overall average.
- “Low entropy proves the rule is invalid.” Entropy describes concentration, not validity. Judge the allocation against the design’s comparison goals and prespecified operating characteristics.
- “A probability floor guarantees precision.” A floor limits how small the chance can be at any one decision; it cannot guarantee enough remaining assignments, outcome completeness, or a precise estimate.
These diagnostics do not replace a full design evaluation. The normal-outcome precision calculation is only an illustration, and adaptive assignment can affect the sampling distribution and analysis in ways that a count-based formula does not capture. The key design lesson is narrower: when an arm remains active but receives near-zero assignment probability, its future data support may be weak, and that weakness should be visible before interpreting comparisons involving the arm.
Check Your Understanding
Answer each question, then compare your reasoning with the solutions below.
- At one update, four active arms have probabilities 0.70, 0.20, 0.09, and 0.01. What does the 0.01 probability tell you, and what additional information is needed to assess practical support?
- An arm has a constant probability of 0.05 over the next 20 assignments. Calculate its expected number of assignments and the probability it receives none.
- Under the independent normal, common-variance model with fixed arm counts, why does a small arm count increase the variance of a two-arm mean difference?
- Does a probability floor of 0.02 guarantee at least two assignments to an arm over the next 100 decisions? Explain.
Solutions. (1) The arm has only a 1% chance at this decision, indicating a concentrated assignment. Assess how often this occurs, how many decisions remain, the arm’s other conditional probabilities, and the resulting expected and realized counts or zero-assignment risk. (2) The expected count is \(20(0.05)=1\). The probability of none is \(0.95^{20}\approx0.3585\). (3) The conditional variance is \(\sigma^2(1/n_a+1/n_b)\); a small \(n_a\) makes \(1/n_a\) large, increasing the variance. (4) No. A floor of 0.02 gives an expected count of at least 2 only if it applies at all 100 decisions, but expectation is not a guarantee of a realized count. Even at constant probability 0.02, the chance of zero assignments is \(0.98^{100}\approx0.1326\), and counts can also be positive but less than two.