Neyman Allocation Across Multiple Treatment Arms

Why Allocation Across Arms Matters

When a trial has several treatment arms, a fixed total sample must be divided among them. If outcome variability differs across arms, equal sample sizes need not give the smallest overall sampling variance for the arm means. A more variable arm generally benefits from receiving more participants: its mean is less precisely estimated per participant, so additional observations can reduce its contribution to the total uncertainty.

There is an important qualification. An allocation cannot generally minimize the variance of every arm mean separately: increasing one arm’s sample size leaves fewer participants for the others. To make “variance-minimizing” precise, this tutorial uses the sum of the arm-mean variances as its objective, giving each treatment-arm mean equal weight. The resulting allocation is the multiarm Neyman target under equal per-participant costs.

Definition: Suppose there are $K$ treatment arms. Arm $i$ has $n_i$ participants, an outcome with variance $\sigma_i^2$, and a sample mean $\bar{x}_i$. Under independent observations within and across arms, $\operatorname{Var}(\bar{x}_i)=\sigma_i^2/n_i$. For fixed total sample size $N$, the equal-weight sum-of-variances objective is $$ V(n_1,\ldots,n_K)=\sum_{i=1}^{K}\frac{\sigma_i^2}{n_i}, \qquad \sum_{i=1}^{K} n_i=N. $$

Assumptions and the Allocation Target

The target applies when the arm outcome variances are treated as known planning quantities and are finite and strictly positive. Within each arm, outcomes are assumed to be independent draws from a distribution with a common arm-specific mean and variance; observations from different arms are also independent. The mean estimator is the ordinary sample mean. These assumptions give the stated variance formula; normal outcome distributions are not required for that formula.

The total $N$ is fixed, each participant has the same cost, and all arm means receive equal weight in the objective. For now, $n_i$ is a continuous target used to define proportions; actual participant counts must be integers. The allocation concerns precision for estimating the separate arm means, not a particular treatment contrast or a welfare objective.

Theorem: Under the assumptions above, the allocation that minimizes $V$ has proportions $$ q_i^*=\frac{\sigma_i}{\sum_{j=1}^{K}\sigma_j}, \qquad n_i^*=Nq_i^*=\frac{N\sigma_i}{\sum_{j=1}^{K}\sigma_j}. $$ The minimum sum of variances is $$ V_{\min}=\frac{\left(\sum_{i=1}^{K}\sigma_i\right)^2}{N}. $$

The standard deviation $\sigma_i$, rather than the variance $\sigma_i^2$, appears in the allocation share. A large variance does increase the contribution $\sigma_i^2/n_i$ to the objective, but taking the square-root relationship into account when allocating the fixed total gives shares proportional to standard deviations.

Why the Target Minimizes the Objective

A short inequality establishes both the minimum and its equality condition. Apply the Cauchy–Schwarz inequality to the two $K$-component vectors with entries $\sigma_i/\sqrt{n_i}$ and $\sqrt{n_i}$. Because $\sum_i n_i=N$,

\left(\sum_{i=1}^{K}\sigma_i\right)^2 = \left(\sum_{i=1}^{K}\frac{\sigma_i}{\sqrt{n_i}}\sqrt{n_i}\right)^2 \leq \left(\sum_{i=1}^{K}\frac{\sigma_i^2}{n_i}\right) \left(\sum_{i=1}^{K}n_i\right) = V N.

Dividing by the positive total $N$ shows that every feasible allocation has $V\geq(\sum_i\sigma_i)^2/N$. Equality in Cauchy–Schwarz holds exactly when the two vectors are proportional. Thus there is a constant $c>0$ such that $\sigma_i/\sqrt{n_i}=c\sqrt{n_i}$ for every arm, or $n_i=\sigma_i/c$. The constraint $\sum_i n_i=N$ then gives $c=(\sum_i\sigma_i)/N$, and hence $n_i=N\sigma_i/\sum_j\sigma_j$. Substitution into $V$ gives the stated minimum.

Worked Allocations

Worked Example: Three Arms With Different Variability

Consider a hypothetical three-arm trial with a fixed total of $N=180$ participants. Suppose the planning standard deviations are $4$, $6$, and $8$ outcome units in arms 1, 2, and 3, respectively. Assume independent outcomes, equal costs, positive finite variances, and that the target is to minimize the equal-weight sum of the three mean variances.

The sum of standard deviations is $4+6+8=18$. Therefore the target shares and sample sizes are $$ (q_1^*,q_2^*,q_3^*)= \left(\frac{4}{18},\frac{6}{18},\frac{8}{18}\right) = \left(\frac{2}{9},\frac{1}{3},\frac{4}{9}\right), \qquad (n_1^*,n_2^*,n_3^*)=(40,60,80). $$

The arm-mean variances at these targets are $4^2/40=0.4000$, $6^2/60=0.6000$, and $8^2/80=0.8000$ outcome-units squared. Their sum is $0.4000+0.6000+0.8000=1.800$ outcome-units squared. The theorem gives the same result: $$ V_{\min}=\frac{(4+6+8)^2}{180}=\frac{324}{180}=1.800. $$

For comparison, equal allocation gives 60 participants per arm and a sum of variances of $16/60+36/60+64/60=116/60=1.933$ outcome-units squared. Under this stated objective, the Neyman target lowers the sum by $1.933-1.800=0.1333$ outcome-units squared. It does not make every arm mean equally precise; the arm-mean variances differ because their outcome variances differ.

Worked Example: Four Arms and the Standard-Deviation Rule

Suppose a hypothetical study has four arms with planning standard deviations $2$, $3$, $3$, and $6$ outcome units and fixed total $N=140$. Assume the same independence, equal-cost, and equal-weight conditions as in the preceding example.

The sum of standard deviations is $2+3+3+6=14$. The shares are $2/14,3/14,3/14,6/14$, so the continuous targets are $20,30,30,60$ participants. These happen to be integers and sum to 140.

The four contributions to the objective are $$ \frac{2^2}{20}=0.2000,\quad \frac{3^2}{30}=0.3000,\quad \frac{3^2}{30}=0.3000,\quad \frac{6^2}{60}=0.6000. $$ Thus $V=0.2000+0.3000+0.3000+0.6000=1.400$ outcome-units squared. The minimum formula confirms the calculation: $(2+3+3+6)^2/140=196/140=1.400$.

The arm with standard deviation 6 receives three times the participants of either arm with standard deviation 2, because $6/2=3$. Allocating in proportion to the variances instead would assign shares proportional to $4,9,9,36$, which is not the minimizer of the stated objective.

Worked Example: Noninteger Continuous Targets

For a hypothetical three-arm trial, suppose the planning standard deviations are $2$, $4$, and $5$ outcome units and $N=125$. Assume the variances are positive and finite, observations are independent as specified above, and the allocation objective gives equal weight to the three mean variances.

The standard deviations sum to 11, so the target proportions are $2/11$, $4/11$, and $5/11$. Multiplying by the total gives continuous targets $$ (n_1^*,n_2^*,n_3^*)= \left(\frac{250}{11},\frac{500}{11},\frac{625}{11}\right) \approx(22.73,45.45,56.82). $$ Their sum is exactly 125 before rounding. The minimum sum of variances at the continuous target is $$ V_{\min}=\frac{(2+4+5)^2}{125}=\frac{121}{125}=0.9680 $$ outcome-units squared.

This calculation determines the optimal proportions and continuous targets, not an integer implementation. In an actual trial, assigning fractional participants is impossible, and the integer counts will generally produce a value of $V$ slightly different from this continuous minimum.

Interpretation, Diagnostics, and Alternatives

The target is a precision rule: under its assumptions, larger $\sigma_i$ means a larger sample-size share. It does not imply that the arm has a more favorable treatment effect, nor does it establish a causal benefit. In a response-adaptive design, these proportions can be understood as allocation targets; the realized arm counts may differ from the targets because participants arrive sequentially and randomization is stochastic.

The central diagnostic question is whether the quantities and objective used to obtain the target still describe the design problem. Check that the planning standard deviations are expressed on the same outcome scale, are positive, and are plausible for the relevant arm populations. Check that the total sample and equal-cost assumption are appropriate, and that the aim really is to estimate all arm means with equal weight. If any of these conditions fails, the displayed shares need not be optimal for the actual goal.

Allocation or objective	What it targets	Why it may differ
Equal allocation	The same sample size in every arm	It does not use differences in arm outcome variability.
Multiarm Neyman allocation here	Minimum equal-weight sum of arm-mean variances	It assumes equal costs and uses shares proportional to arm standard deviations.
Cost-adjusted allocation	Precision when per-participant costs differ	Unequal costs change the optimization problem and its target.
Allocation for a particular contrast	Precision for a specified comparison	A contrast-specific objective need not give the same shares as an equal-weight objective for all means.

Equal allocation remains a useful benchmark, particularly when variances are similar or operational simplicity matters. The two-arm and binary-outcome cases are special settings for which the same variance-minimization principle can be applied with their relevant outcome variances. But the multiarm formula is not a universal prescription: changing the objective, costs, dependence structure, or allocation constraints can change the answer.

A common mistake is to say, “Arm 3 has four times the variance of arm 1, so it should receive four times as many participants.” The correction is to use standard deviations: if its variance is four times as large, its standard deviation is twice as large, so its Neyman share is twice as large as arm 1’s, relative to the other arms’ shares. Another mistake is to describe this allocation as minimizing every pairwise comparison’s variance. It minimizes the specified sum of arm-mean variances; a different contrast-focused objective can favor a different allocation.

Key takeaway: With fixed total sample size, independent arm outcomes, equal costs, and equal weight on every arm mean, allocate in proportion to the arm-specific standard deviations. The target minimizes the sum of the sample-mean variances, not every possible precision criterion at once.

Check Your Understanding

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

For three arms with standard deviations $3$, $3$, and $6$, what are the target proportions?
Why does the allocation use $\sigma_i$, rather than $\sigma_i^2$, in its numerator?
With $N=120$ and standard deviations $2$, $4$, and $6$, calculate the continuous target sample sizes and the minimum sum of variances.
Does the multiarm Neyman allocation minimize the variance of every pairwise treatment comparison? Explain.
Name two conditions under which this target may not be optimal for the intended design.

Solutions. 1. The standard deviations sum to 12, so the proportions are $3/12,3/12,6/12$, or $1/4,1/4,1/2$. 2. The Cauchy–Schwarz equality condition under $\sum_i n_i=N$ gives $n_i$ proportional to $\sigma_i$; the full variances $\sigma_i^2$ enter the objective, but the optimizing shares are proportional to their square roots. 3. The standard deviations sum to 12, so the targets are $120(2/12)=20$, $120(4/12)=40$, and $120(6/12)=60$. The minimum is $(2+4+6)^2/120=144/120=1.200$ outcome-units squared. 4. No. It minimizes the equal-weight sum of the arm-mean variances; a specified pairwise contrast uses a different objective and may have a different optimal allocation. 5. Examples include unequal per-participant costs, a contrast-specific rather than equal-weight objective, dependent outcomes, or planning variances that do not represent the arm populations.