Extending the Rule to Multiple Treatment Arms

From Two Arms to Multiple Arms

With more than two treatments, a DBCD must decide how to correct allocation across several arms at once. The ingredients are familiar: an estimated target share for each arm, the current proportion of patients assigned to each arm, and a tuning parameter that controls the strength of correction. The result is a probability for assigning the next patient to each arm. These three quantities have distinct roles: the target is a desired allocation pattern, the current proportions describe allocations already made, and the next-patient probabilities govern the next random assignment.

The rule below extends the two-arm ratio correction to $K$ treatment arms. It uses target shares estimated from information available before the next assignment. The estimation method is not part of the rule itself; for example, a model-based procedure may supply the estimates, but the allocation step takes those estimates as inputs.

Definition: Consider $K$ treatment arms, indexed by $k=1,\ldots,K$. After $n$ patients have been assigned, let $N_{k,n}$ be the number assigned to arm $k$, and let $\hat{\pi}_{k,n}=N_{k,n}/n$ be its current allocation proportion, for $n>0$. Let $\hat{\rho}_{k,n}$ be the estimated target share for arm $k$, calculated using information available before patient $n+1$ is assigned. The estimated shares satisfy $\sum_{k=1}^K\hat{\rho}_{k,n}=1$. For tuning parameter $\gamma\geq 0$, the multi-arm DBCD next-patient probability is $$ q_{k,n+1} = \frac{\hat{\rho}_{k,n} \left(\hat{\rho}_{k,n}/\hat{\pi}_{k,n}\right)^\gamma} {\displaystyle\sum_{j=1}^K \hat{\rho}_{j,n} \left(\hat{\rho}_{j,n}/\hat{\pi}_{j,n}\right)^\gamma}, \qquad k=1,\ldots,K. $$ Conditional on the trial history before assignment, patient $n+1$ is assigned to arm $k$ with probability $q_{k,n+1}$.

For this form of the rule, assume each current allocation proportion and each estimated target share is positive, and that $\gamma$ is finite. Positivity makes each ratio in the formula well-defined. The target estimates and current counts must be available before the next assignment; using the next patient’s outcome to set that patient’s treatment probability would violate this sequencing. The assignment is randomized according to the displayed probabilities, rather than chosen deterministically.

When $n=0$, current allocation proportions cannot be computed as $N_{k,n}/n$. Likewise, an arm with no assignments has current proportion zero. A trial therefore needs an initialization procedure that places patients on every arm before applying this form of the rule, or another prespecified method for handling zero counts. Initialization is not a cosmetic detail: without a positive current proportion, the ratio correction is undefined.

How the Correction Works Across Arms

Each arm’s unnormalized contribution to the assignment probabilities is its target share multiplied by a correction factor. If an arm’s estimated target share is large relative to its current allocation proportion, the ratio $\hat{\rho}_{k,n}/\hat{\pi}_{k,n}$ exceeds one, and the correction increases that arm’s relative weight. If its current proportion is large relative to its target share, the ratio is below one, and the correction reduces its relative weight. The parameter $\gamma$ controls how strongly that ratio affects the next assignment.

A short derivation makes the multi-arm comparison precise. Write $w_{k,n}=\hat{\rho}_{k,n}(\hat{\rho}_{k,n}/\hat{\pi}_{k,n})^\gamma$ for the unnormalized weight in the rule. Since all next-patient probabilities share the same denominator, the ratio of two arms’ probabilities equals the ratio of their weights:

\frac{q_{i,n+1}}{q_{j,n+1}} = \frac{w_{i,n}}{w_{j,n}} = \frac{\hat{\rho}_{i,n}}{\hat{\rho}_{j,n}} \left( \frac{\hat{\rho}_{i,n}/\hat{\pi}_{i,n}} {\hat{\rho}_{j,n}/\hat{\pi}_{j,n}} \right)^\gamma.

Thus the target-share odds between arms $i$ and $j$ are multiplied by a correction that compares their target-to-current ratios. If arm $i$ is more underrepresented relative to its target than arm $j$, its correction factor increases its assignment probability relative to arm $j$. This is a relative statement: because the probabilities must sum to one, a change in one arm’s probability also affects the probabilities of the other arms.

At $\gamma=0$, the correction factors equal one, so the next-patient probabilities equal the estimated target shares. For positive $\gamma$, current allocation proportions influence the probabilities; larger values strengthen their influence. This description does not imply that every arm’s probability moves monotonically toward its target in every possible multi-arm state. The arms share a denominator, and the rule responds to the whole allocation vector.

Worked Example: Three Arms With Unequal Current Allocation

Suppose a hypothetical trial has three arms. Before the next assignment, 20 of 40 patients have been assigned to arm 1, 10 to arm 2, and 10 to arm 3. The estimated target shares are $0.50$, $0.30$, and $0.20$, respectively, and $\gamma=1$. All target shares and current proportions are positive, and the shares and proportions each sum to one, so the rule’s stated conditions hold.

The current proportions are $\hat{\pi}_{1,n}=20/40=0.50$, $\hat{\pi}_{2,n}=10/40=0.25$, and $\hat{\pi}_{3,n}=10/40=0.25$. With $\gamma=1$, the three unnormalized weights are:

w_{1,n}=0.50(0.50/0.50)=0.50,\qquad w_{2,n}=0.30(0.30/0.25)=0.36,\qquad w_{3,n}=0.20(0.20/0.25)=0.16.

Their sum is $0.50+0.36+0.16=1.02$. Dividing each weight by this sum gives:

(q_{1,n+1},q_{2,n+1},q_{3,n+1}) = (0.50/1.02,\;0.36/1.02,\;0.16/1.02) \approx(0.4902,\;0.3529,\;0.1569).

The probabilities sum to one, up to rounding. Arm 2 is below its target share in the current allocations, so its next-patient probability exceeds its target share. These are probabilities for one assignment, not guarantees about the treatment proportions after that assignment.

Worked Calculations With Estimated Targets

The estimated target shares can change as response information accumulates. The allocation formula is applied using the current estimates at each assignment; it does not require those estimates to remain fixed. In the next example, the target estimates are treated as prespecified inputs for the single allocation calculation. The example does not specify or evaluate the procedure that produced them.

Worked Example: Updating From a New Set of Target Estimates

In a hypothetical three-arm trial, the current counts are $12$, $14$, and $14$, for a total of 40 patients. The current target estimates are $0.40$, $0.35$, and $0.25$, and $\gamma=0.5$. The target estimates are positive and sum to one; the current proportions are also positive and sum to one.

The current proportions are $0.30$, $0.35$, and $0.35$. The weights use the square root of each target-to-current ratio:

\begin{aligned} w_{1,n}&=0.40(0.40/0.30)^{0.5}\approx0.4619,\\ w_{2,n}&=0.35(0.35/0.35)^{0.5}=0.3500,\\ w_{3,n}&=0.25(0.25/0.35)^{0.5}\approx0.2113. \end{aligned}

The weight sum is approximately $0.4619+0.3500+0.2113=1.0232$. Normalizing gives:

(q_{1,n+1},q_{2,n+1},q_{3,n+1}) \approx (0.4619/1.0232,\;0.3500/1.0232,\;0.2113/1.0232) \approx(0.4514,\;0.3421,\;0.2065).

Although arm 2 is currently at its estimated target share, its next-patient probability is not exactly $0.35$. The other arms’ corrections change the common normalizing denominator. The result is a probability vector for the next patient, not a set of separately applied adjustments to the observed counts.

What the Rule Does Not Guarantee

The probabilities govern random assignments one patient at a time. To see the distinction from a realized allocation proportion, let $X_{k,n+1}$ equal 1 if patient $n+1$ is assigned to arm $k$, and 0 otherwise. Conditional on the trial history, $E(X_{k,n+1})=q_{k,n+1}$. Since the new allocation proportion is $(N_{k,n}+X_{k,n+1})/(n+1)$, its conditional expectation is:

E(\hat{\pi}_{k,n+1}\mid\text{trial history}) = \frac{n\hat{\pi}_{k,n}+q_{k,n+1}}{n+1}, \qquad E(\hat{\pi}_{k,n+1}\mid\text{trial history})-\hat{\pi}_{k,n} = \frac{q_{k,n+1}-\hat{\pi}_{k,n}}{n+1}.

This identity shows that the expected change in an arm’s allocation proportion depends on the difference between its next-patient probability and its current proportion. It does not say that the next realized assignment will produce that expected change, or that the current proportions will equal their targets after one step. Random variation remains, and the target estimates may also change as new outcomes arrive.

The rule also does not, by itself, establish that the estimated target is ethically or statistically appropriate, that the response estimates are unbiased, or that the allocation proportions converge to a fixed value. Those conclusions require assumptions and arguments about the target-estimation procedure, the response process, and the sequential design. Here the focus is the allocation calculation conditional on the current estimates and counts.

Worked Example: A Strong Correction From Equal Current Proportions

Suppose a hypothetical trial has three arms with 10 patients assigned to each, so the current proportions are all $1/3$. The estimated target shares are $0.60$, $0.30$, and $0.10$, and $\gamma=1$. Each current proportion and target share is positive, and both sets sum to one.

The weights are:

w_{1,n}=0.60(0.60/(1/3))=1.08,\qquad w_{2,n}=0.30(0.30/(1/3))=0.27,\qquad w_{3,n}=0.10(0.10/(1/3))=0.03.

The sum is $1.08+0.27+0.03=1.38$, so the next-patient probabilities are:

(q_{1,n+1},q_{2,n+1},q_{3,n+1}) = (1.08/1.38,\;0.27/1.38,\;0.03/1.38) \approx(0.7826,\;0.1957,\;0.0217).

The probability for arm 1 is substantially above its current proportion because its estimated target share is much larger. The calculation remains randomized: even with a next-patient probability of approximately $0.7826$, assignment to arm 1 is not certain.

Comparisons, Diagnostics, and Failure Modes

At $\gamma=0$, the rule assigns the next patient according to the estimated target shares alone. For $\gamma>0$, it also responds to the current allocation proportions. A fixed-probability design is different in that its assignment probabilities need not respond to either updated target estimates or current allocation. Deterministic assignment to the arm with the largest target deficit is different again: it removes the randomization represented by the probabilities. The DBCD retains randomization while using feedback to adjust assignments.

Several checks can catch implementation errors before they affect a trial:

Check the inputs. Confirm that the estimated target shares are nonnegative and sum to one, and that all shares and current proportions used in this formula are positive.
Check the timing. Confirm that counts and target estimates use only information available before the patient is assigned.
Check the output. Each calculated probability should be positive under the stated conditions, and their sum should equal one apart from rounding.
Check the scale of correction. A very large $\gamma$ can concentrate assignment probability on arms whose target-to-current ratios are relatively high. Inspect probabilities for unintended near-zero values or numerical underflow.
Check the distinction between probability and count. A next-patient probability is not an observed allocation proportion, and the observed proportion can differ from its expected value in a finite sample.

A frequent mistake is to say, “Arm 2 has a target share of 0.30, so the next patient must have a 0.30 probability of receiving arm 2.” That is true in the special case $\gamma=0$, but not generally. For positive $\gamma$, the current proportion affects the arm’s weight, and the common denominator also depends on every other arm. Another mistake is to apply the formula when an arm has zero assignments: its current proportion is zero, so the ratio is undefined. The correction is to use a prespecified initialization or a separately defined zero-count procedure, not to divide by zero or silently replace the count.

Key takeaway: A multi-arm DBCD combines each arm’s estimated target share with its target-to-current allocation ratio, then uses the resulting relative weights to randomize the next patient. The probabilities adapt assignment; they do not force realized counts to match targets at each step.

Check Your Understanding

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

What is the difference between $\hat{\rho}_{k,n}$, $\hat{\pi}_{k,n}$, and $q_{k,n+1}$?
For $\gamma=0$, what is $q_{k,n+1}$?
A two-arm trial has current proportions $0.60$ and $0.40$, target shares $0.40$ and $0.60$, and $\gamma=1$. Calculate the next-patient probabilities.
Why is the displayed rule not directly defined if an arm has no patients assigned?
Does a next-patient probability of $0.70$ mean that exactly 70% of patients will have been assigned to that arm after the next assignment? Explain.

Solutions. (1) $\hat{\rho}_{k,n}$ is arm $k$’s estimated target share; $\hat{\pi}_{k,n}$ is its observed share among the $n$ assignments already made; $q_{k,n+1}$ is the conditional probability of assigning the next patient to that arm. (2) When $\gamma=0$, every correction factor is one, and the weights sum to one, so $q_{k,n+1}=\hat{\rho}_{k,n}$. (3) The weights are $0.40(0.40/0.60)=0.2667$ and $0.60(0.60/0.40)=0.9000$, with sum $1.1667$. Thus the probabilities are approximately $0.2286$ and $0.7714$. (4) Its current proportion is zero, making the target-to-current ratio undefined; the design needs a prespecified initialization or another defined procedure. (5) No. The probability describes the next randomized assignment, while the realized proportion after that assignment depends on the previous count and the actual assignment outcome.