Normalizing Multi-Arm Allocation Weights

Why Normalization Matters

With multiple treatment arms, the allocation rule must turn information about all arms into a single set of probabilities for the next participant. A target share says what long-run allocation the design aims to track; the current shares describe the allocation so far. The DBCD rule uses their ratios to adjust the next assignment probabilities. Because every arm’s probability depends on the other arms’ weights, those adjusted weights must be normalized jointly.

This tutorial treats the target allocation vector as given at the current assignment. It focuses on the calculation that converts that vector and the current allocation shares into probabilities. It does not require a particular method for constructing the target from response estimates.

Shares, Ratios, and Weights

Suppose there are \(K\) treatment arms. Before participant \(n+1\) is assigned, let \(N_k(n)\) be the number assigned to arm \(k\), and let \(n=\sum_{k=1}^K N_k(n)\) be the number already assigned. The current allocation share of arm \(k\) is \(x_k(n)=N_k(n)/n\). Let \(\pi_k\) be its target share, with \(\sum_{k=1}^K \pi_k=1\). The tuning parameter is \(\gamma\geq 0\).

Writing \(p_k(n+1)\) for the probability of assigning participant \(n+1\) to arm \(k\), the rule is

Equivalently, define the unnormalized weight \(w_k(n)=\pi_k\{\pi_k/x_k(n)\}^{\gamma}\), and let \(W(n)=\sum_{j=1}^K w_j(n)\). Then \(p_k(n+1)=w_k(n)/W(n)\). The ratios compare shares and are dimensionless. If a target share exceeds the current share, its ratio exceeds one; if it falls short, its ratio is below one. The power \(\gamma\) controls how strongly that ratio changes the weight, while the factor \(\pi_k\) retains the target share as the baseline weight.

Why These Weights Are Normalized This Way

The multi-arm extension needs both a correction and a baseline. The correction is the powered ratio \(\{\pi_k/x_k(n)\}^{\gamma}\); the baseline is \(\pi_k\). A natural construction is to multiply these arm-specific positive quantities to obtain weights, then divide each weight by their common total. The resulting probabilities are nonnegative and collectively describe one assignment.

This is the normalization check: it follows directly from using the same total weight in every denominator. Under the stated conditions, every \(w_k(n)>0\) and \(W(n)>0\), so each probability is positive as well. At \(\gamma=0\), the powered ratios equal one and the probabilities reduce to the target shares. If current shares equal target shares, every ratio is one, so the probabilities also equal the target shares for any \(\gamma\).

These checks are useful for distinguishing the rule from a ratio-only formula. Normalizing \(\{\pi_k/x_k(n)\}^{\gamma}\) without the baseline factor \(\pi_k\) would give equal probabilities when \(\gamma=0\), not the target allocation in general. That is a different rule. The DBCD weights here preserve the target allocation as the \(\gamma=0\) case.

Derivation and Checks in Practice

The normalization step can be justified without assuming that the raw weights already sum to one. Start with positive weights \(w_1,\ldots,w_K\). A probability vector proportional to these weights has the form \(p_k=cw_k\), where \(c\) is a common positive constant. Requiring the probabilities to sum to one gives \(c\sum_j w_j=1\), hence \(c=1/\sum_j w_j\). Substitution gives \(p_k=w_k/\sum_jw_j\). Thus the common denominator is required by the probability constraint; it is not an optional adjustment.

Before using the calculation, check the inputs and output. The current counts should sum to \(n\); the target shares should sum to one; and each share used in a denominator should be positive. After computing the weights, their total should be positive. Finally, calculate all probabilities with the same total weight and verify that they sum to one before rounding. A sum slightly different from one after rounding can be harmless; a material discrepancy usually signals an arithmetic or transcription error.

Failure Modes and Alternatives

The formula as stated requires positive current and target shares. If an arm has \(N_k(n)=0\), its current share is zero and \(\pi_k/x_k(n)\) is undefined. A design needs a suitable initialization or another explicitly specified boundary rule before applying this formula; silently substituting zero, an arbitrary small value, or an infinite weight does not follow from the rule. A zero target share also falls outside the positive-target condition used here and may require a separate design definition. These cases should be handled deliberately, not by pretending the ordinary calculation still applies.

The principal alternative at a given state is to randomize directly with probabilities equal to the target shares, \(p_k(n+1)=\pi_k\), without correcting for current allocation shares. That is simple and remains well-defined for positive target shares, but it does not use the current imbalance in this assignment step. The normalized DBCD rule incorporates that imbalance through the ratios. Neither calculation changes the target vector itself; the key distinction is whether current shares are used to adjust the next assignment probabilities.

Another common error is to compute weights and report them as probabilities. Weights such as \((0.625,0.225,0.200)\) are not probabilities unless they sum to one. Divide every weight by the same total. Also avoid rounding ratios or weights too early: small rounding errors across several arms can obscure whether normalization is correct.

Normalization guarantees a valid probability vector under the stated conditions; it does not guarantee that the allocation will equal the target at each step, nor does it remove uncertainty in the realized sequence of assignments. The formula is also sensitive to its inputs: small current shares can produce large ratios, especially when \(\gamma\) is large. Thus, verify the computed vector and its assumptions rather than treating successful normalization as evidence that every design choice is appropriate.