Combining Allocation Adaptation and Arm Dropping

Why the Order of Operations Matters

In a response-adaptive multiarm trial, allocation probabilities depend on accumulating outcome information. At an interim, the same information may also trigger a decision to drop one or more arms. These operations must be specified together: an allocation rule that still assigns probability to an arm that has been dropped is not a valid rule for the next allocation.

A clear procedure separates the calculations. First use the interim data to calculate the allocation scores under the stated adaptation rule. Next apply the arm-dropping rule to identify which arms remain eligible. Finally set the dropped arms’ allocation weights to zero and renormalize the retained arms’ weights so the probabilities sum to one. This makes explicit that allocation probabilities for the next assignments are conditional on the retained set.

Definition: Let $A$ be the set of arms active immediately before an interim decision, and let $R\subseteq A$ be the set retained by the arm-dropping rule. Let $w_j\geq 0$ be the allocation score for arm $j$, calculated from the information available under the prespecified adaptation rule. If $\sum_{k\in R}w_k>0$, the post-dropping allocation probability for a retained arm $j\in R$ is $$ \pi_j^{(R)}=\frac{w_j}{\sum_{k\in R}w_k}. $$ For every dropped arm $j\in A\setminus R$, set $\pi_j^{(R)}=0$. Thus $\sum_{j\in A}\pi_j^{(R)}=1$.

The scores $w_j$ are not necessarily probabilities. For example, they may be posterior mean outcomes, posterior probabilities of being best, or nonnegative functions of posterior uncertainty. The protocol must say which score is used and how it is computed. If an allocation algorithm first converts scores into probabilities $q_j$ over all currently active arms, renormalization is equivalent to $\pi_j^{(R)}=q_j/\sum_{k\in R}q_k$ for retained arms. The two representations produce the same probabilities because the common initial normalization cancels.

A Prespecified Sequence

The following sequence prevents an arm-dropping decision and an allocation update from being applied inconsistently. It assumes that the interim analysis uses the prespecified data cutoff, that the adaptation and dropping rules are defined before the trial, and that the retained set includes at least one arm with a positive allocation score. Bayesian posterior calculations must also follow the specified outcome model and prior. When outcome data are delayed, the available information and its use must follow the trial’s stated interim rules.

Freeze the interim information.
Specify which observations and outcomes are available at this decision. Do not change the data cutoff to favor an allocation or dropping result.

Calculate adaptation scores.
Apply the prespecified response-adaptive rule to the interim information for the arms active before the decision. Record the scores before dropping arms.

Apply the dropping rule.
Evaluate the prespecified criterion for each arm and record the retained set $R$. Apply any control-retention constraint as part of this decision.

Zero dropped-arm weights and renormalize.
Use only the retained arms’ scores in the denominator. Check that the resulting probabilities sum to one and satisfy any allocation constraints.

Use the resulting rule for subsequent assignments.
Document the data, decisions, scores, retained set, and probabilities. The timing and implementation of later allocations must follow the protocol.

The denominator is the sum of scores for retained arms, not the sum for all arms before dropping. Renormalizing first and then merely setting dropped-arm probabilities to zero leaves total probability below one. Recomputing scores after dropping is a different design choice: it can change the relative probabilities among retained arms and must not be substituted silently for the specified rule.

Worked Example: Four Arms With Posterior-Mean Response Weights

Hypothetical data. A trial has control $C$ and three experimental arms $A$, $B$, and $D$. The binary outcome is response by the prespecified assessment time. Assume independent Bernoulli outcomes within each arm and independent $\mathrm{Beta}(1,1)$ priors for the response probabilities. At an interim, the completed data are $15/30$ responses for $C$, $18/30$ for $A$, $8/30$ for $B$, and $4/30$ for $D$. For this example, the adaptation score is the posterior mean response probability, and the dropping rule drops an experimental arm if its posterior mean is below $0.20$. Control is retained by design.

With a $\mathrm{Beta}(1,1)$ prior and $r$ responses among $n$ participants, the posterior is $\mathrm{Beta}(1+r,1+n-r)$, whose mean is $(1+r)/(2+n)$. The scores are therefore

w_C=\frac{16}{32}=0.5000,\quad w_A=\frac{19}{32}=0.5938,\quad w_B=\frac{9}{32}=0.2813,\quad w_D=\frac{5}{32}=0.1563.

The dropping threshold is applied to the experimental arms’ scores. Arms $A$ and $B$ are retained because $0.5938\geq0.20$ and $0.2813\geq0.20$. Arm $D$ is dropped because $0.1563<0.20$. The retained set is $R=\{C,A,B\}$, and its total score is

\sum_{k\in R}w_k=0.5000+0.5938+0.2813=1.3751 \quad\text{(using unrounded scores, this is }44/32=1.3750\text{).}

Using the exact fractions avoids a rounding discrepancy. The resulting probabilities are

\pi_C^{(R)}=\frac{16}{44}=0.3636,\quad \pi_A^{(R)}=\frac{19}{44}=0.4318,\quad \pi_B^{(R)}=\frac{9}{44}=0.2045,\quad \pi_D^{(R)}=0.

The retained probabilities sum to $44/44=1$. These are the probabilities for subsequent assignments under this example’s rule, conditional on the retained set. They are not probabilities that the arms are effective, nor do they establish that allocation to an arm causes a better response. The calculation depends on the specified binary-outcome model, priors, completed interim data, posterior-mean score, and threshold; changing any of these can change the result.

Why Renormalization Works

Suppose the pre-dropping allocation probabilities are $q_j=w_j/\sum_{k\in A}w_k$, with positive total score over $A$. Once the trial restricts assignment to $R$, the probability assigned to a retained arm should be proportional to its original score among the eligible arms. Conditioning the original allocation distribution on being in $R$ gives

\frac{q_j}{\sum_{k\in R}q_k} = \frac{w_j/\sum_{k\in A}w_k} {\sum_{k\in R}w_k/\sum_{k\in A}w_k} = \frac{w_j}{\sum_{k\in R}w_k}, \qquad j\in R.

This establishes both the renormalization formula and its interpretation: among the arms still eligible, the original relative weights are preserved. The result is a probability distribution because

\sum_{j\in R}\pi_j^{(R)} = \sum_{j\in R}\frac{w_j}{\sum_{k\in R}w_k} = \frac{\sum_{j\in R}w_j}{\sum_{k\in R}w_k} =1.

This derivation assumes that the scores are nonnegative and that the retained arms have a positive score total. If either condition fails, the formula does not define valid allocation probabilities. The protocol therefore needs a fallback rule for exceptional cases rather than an improvised denominator.

Worked Example: Posterior Probabilities of Being Best

Hypothetical rule output. Suppose a four-arm trial, including control $C$, uses posterior probabilities of being best as its nonnegative allocation weights. At an interim, the prespecified posterior calculation gives $w_C=0.10$, $w_A=0.35$, $w_B=0.30$, and $w_D=0.25$. These weights sum to one. An independently specified futility rule drops $D$, while $C$, $A$, and $B$ are retained.

The retained weight total is $0.10+0.35+0.30=0.75$. Renormalizing only those weights gives

\pi_C^{(R)}=\frac{0.10}{0.75}=0.1333,\quad \pi_A^{(R)}=\frac{0.35}{0.75}=0.4667,\quad \pi_B^{(R)}=\frac{0.30}{0.75}=0.4000,\quad \pi_D^{(R)}=0.

The probabilities sum to $1.0000$. Simply setting $\pi_D$ to zero while leaving the other probabilities unchanged would sum to $0.75$, so it would not define a randomization distribution. This example treats the posterior weights as given outputs of the stated Bayesian model; it does not imply that posterior probability of being best is the only appropriate adaptation score.

Control Constraints and Alternative Choices

A protocol may require a minimum allocation probability for control, for example to preserve concurrent control information. Ordinary renormalization does not guarantee that requirement. If the initial post-renormalization control probability is below a prespecified floor $c_{\min}$, a constrained rule can assign $c_{\min}$ to control and divide the remaining probability among retained experimental arms in proportion to their scores. This requires at least one retained experimental arm with positive total score and $0\leq c_{\min}<1$. The exact constraint-handling method must be specified in advance.

Worked Example: Applying a Minimum Control Probability

Return to the retained scores from the first example: $w_C=16/32$, $w_A=19/32$, and $w_B=9/32$. Suppose the protocol requires control probability of at least $0.40$. Ordinary renormalization gives $\pi_C=16/44=0.3636$, which is below the floor. Set $\pi_C=0.40$. The remaining $0.60$ is divided between $A$ and $B$ in proportion to $19:9$:

\pi_A=0.60\left(\frac{19}{19+9}\right)=0.4071,\qquad \pi_B=0.60\left(\frac{9}{19+9}\right)=0.1929.

The probabilities are $0.4000+0.4071+0.1929=1.0000$, and the control floor is met. This constrained result differs from simple renormalization, so the protocol must state whether the floor is applied and how it modifies the allocation rule.

Two common approaches should not be conflated. One approach calculates weights from all active arms and then conditions those weights on the retained set, preserving the retained arms’ relative weights. Another approach recalculates the adaptation scores using a model or rule that itself changes after an arm is removed. The second may be defensible, but it does not generally yield the same probabilities; it requires a separately specified recalculation rule. Likewise, dropping an arm and then redistributing its probability equally is not equivalent to score-based renormalization unless the retained scores happen to be equal.

Diagnostics, Failure Modes, and Limitations

At each interim, check that the retained set is consistent with the dropping criterion, that dropped arms receive zero probability, and that retained probabilities sum to one within the chosen numerical tolerance. Also check control-retention and minimum-allocation constraints. Keep adequate numerical precision until the final probabilities are reported: rounding intermediate scores can make a displayed sum differ slightly from one.

Important failure modes include a zero total retained score, all experimental arms meeting a dropping rule when the design requires at least one experimental arm to continue, and incompatible constraints—for example, a control floor that leaves no probability for any required retained experimental arm. These situations need prespecified fallback actions, such as reverting to fixed probabilities among eligible arms or pausing for a review. If no retained arm has a positive score, the renormalization formula is undefined; assigning arbitrary probabilities after observing that fact can alter the design.

A further diagnostic is whether the scores and the dropping rule use compatible information and timing. An arm should not be evaluated using one interim data cutoff for dropping and another for allocation unless that difference is explicitly part of the design. Review the allocation path and the recorded interim calculations for abrupt changes that could indicate a coding or implementation error. A valid probability calculation alone does not establish desirable operating characteristics: the combined design should be evaluated for patient allocation, estimation, and decision performance under plausible trial scenarios.

Wrong: “Arm $D$ has been dropped, so I can set its probability to zero and keep the other probabilities unchanged.” Correction: The remaining probabilities then sum to less than one; renormalize the retained weights.
Wrong: “The allocation probabilities must be recalculated from scratch after dropping, even if the protocol specifies fixed interim scores.” Correction: Apply the specified rule. Recalculation is a different design choice, not an automatic consequence of dropping.
Wrong: “If the control arm is retained, it automatically receives the required minimum share.” Correction: Retention gives control eligibility, not a minimum probability; implement a stated allocation constraint if one is required.

Key takeaway: Calculate scores using the prespecified interim information, apply the dropping rule, and then normalize over the retained arms. Verify a positive denominator, a total probability of one, and all design constraints before using the allocation distribution.

Check Your Understanding

Answer each question using the stated order of operations. Short solutions follow the questions.

Three retained arms have scores $2$, $3$, and $5$. What are their normalized probabilities?
Pre-dropping probabilities are $0.20$, $0.50$, and $0.30$. The third arm is dropped. What are the probabilities among the retained arms?
Why is setting a dropped arm’s probability to zero without changing the other probabilities generally invalid?
In the first worked example, what would happen if the control floor were $0.40$, and how would the experimental probabilities be divided?
What must a protocol specify if all retained arms have allocation score zero?

Solutions. (1) The total score is $2+3+5=10$, so the probabilities are $0.20$, $0.30$, and $0.50$. (2) The retained total is $0.20+0.50=0.70$, giving $0.20/0.70=0.2857$ and $0.50/0.70=0.7143$; the dropped arm receives zero. (3) The remaining probabilities sum to less than one, so they do not form a probability distribution for assignment. (4) Ordinary renormalization gives control $16/44=0.3636$, below the floor; set control to $0.40$, then allocate $0.60$ across $A$ and $B$ in the score ratio $19:9$, giving approximately $0.4071$ and $0.1929$. (5) It must specify a fallback, because division by a zero total score is undefined.