Allocating Participants Across Groups

Why Allocation Matters

A study with a fixed number of participants still has a design choice to make: how many participants should be assigned to each group? A balanced design gives every group the same number. An unbalanced design gives some groups more. This choice affects the precision of estimated group differences and can affect the power of the one-way F test.

Equal allocation is a useful default when groups have similar variability, participation costs are the same, and no particular group comparison should receive priority. It is not a rule that guarantees the greatest power for every possible pattern of group means. To understand both the advantage and the limitation, begin with the model used for the classical one-way ANOVA test.

Model and assumptions: In group $i$, the $j$th outcome is $X_{ij}=\mu_i+\epsilon_{ij}$, for $i=1,\ldots,k$ and $j=1,\ldots,n_i$. For the classical one-way F test, assume the errors are independent and normally distributed with mean zero and a common variance: $\epsilon_{ij}\sim N(0,\sigma^2)$. Thus observations are independent, each group has a normal outcome distribution, and all groups share the same population variance. Let $N=\sum_{i=1}^{k}n_i$ be the fixed total sample size.

The hypotheses for the omnibus test are $H_0:\mu_1=\cdots=\mu_k$ against $H_a:$ not all group means are equal. Under the stated model, the test statistic has an F distribution under $H_0$ with $k-1$ and $N-k$ degrees of freedom. For a specified alternative, its noncentrality parameter is

\lambda=\frac{\sum_{i=1}^{k}n_i(\mu_i-\bar{\mu}_n)^2}{\sigma^2}, \qquad \bar{\mu}_n=\frac{\sum_{i=1}^{k}n_i\mu_i}{N}.

Here $\bar{\mu}_n$ is the allocation-weighted average of the population means. At a fixed significance level $\alpha$, power is the probability that the F statistic exceeds the critical value under the alternative. For fixed $k$, $N$, and $\alpha$, the degrees of freedom and critical value stay the same across allocations; the noncentrality parameter can change. Larger $\lambda$ means greater power for that specified alternative.

Why Equal Allocation Is Efficient Under Symmetry

With equal variances, the variance of the difference between two independent group sample means is

\operatorname{Var}(\bar{X}_i-\bar{X}_j) =\sigma^2\left(\frac{1}{n_i}+\frac{1}{n_j}\right).

If no pair of groups deserves more precision than another, balanced allocation treats every pair symmetrically. In fact, with a fixed total $N$, it minimizes the sum—and therefore the average—of these pairwise variances. This is a precise sense in which equal allocation is efficient. It does not say that it must maximize omnibus power for every specified pattern of means.

Result: For $k$ groups with common variance and fixed $N$, the sum of the variances of all pairwise differences between sample means is minimized when the group sizes are equal, allowing for integer rounding when $N$ is not divisible by $k$.

To see why, each $1/n_i$ appears in exactly $k-1$ pairwise variances. Also, for positive group sizes, the Cauchy–Schwarz inequality gives $(\sum_i n_i)(\sum_i 1/n_i)\geq k^2$. Combining these facts gives

\sum_{i<j}\operatorname{Var}(\bar{X}_i-\bar{X}_j) =\sigma^2(k-1)\sum_{i=1}^{k}\frac{1}{n_i} \geq \sigma^2(k-1)\frac{k^2}{N}.

Equality holds when all $n_i=N/k$. If equal integer group sizes are impossible, sizes differing by at most one are the closest balanced allocation. This result explains the appeal of balance when variances and costs are alike and the study has no special priority among groups.

Worked Example: Comparing Balanced and Unbalanced Precision

Suppose a hypothetical study compares three groups, has $N=30$ participants, and plans for a common standard deviation of 6 outcome units. Compare allocations of $(10,10,10)$ and $(6,10,14)$ using the average variance of the three pairwise differences in sample means. Assume independent observations and the same variance in every group.

With balanced groups, each pairwise variance is $6^2(1/10+1/10)=7.2$ outcome-units squared. The average of the three pairwise variances is therefore $7.2$.

With group sizes $(6,10,14)$, the three variances are $36(1/6+1/10)=9.6$, $36(1/6+1/14)\approx8.571$, and $36(1/10+1/14)\approx6.171$ outcome-units squared. Their average is $(9.6+8.571+6.171)/3\approx8.114$ outcome-units squared.

The balanced allocation has the lower average pairwise variance in this comparison. The unbalanced allocation does give the largest group a more precise sample mean, but it does so at the expense of precision for differences involving the group with only six participants. This calculation compares precision; it does not by itself calculate omnibus power.

How Unequal Variances Change Priorities

If group variances differ, a group with more variable outcomes may need more participants to achieve similar precision. For a particular difference between groups $i$ and $j$, the relevant variance is $\sigma_i^2/n_i+\sigma_j^2/n_j$. If participation costs are equal and the total assigned to those two groups is fixed, this variance is minimized by allocating in proportion to the groups’ standard deviations.

This is a precision rule for a specified difference, not a universal optimum for every omnibus test. It is also important not to quietly retain the classical pooled-variance F test when the common-variance assumption is implausible. Allocation and analysis must be considered together: a design with unequal variances needs an inferential method whose assumptions match those variances.

Worked Example: Allocating More Participants to a More Variable Group

For a hypothetical comparison of groups A and B, suppose planning estimates are $\sigma_A=4$ and $\sigma_B=8$ outcome units. There are 36 participants to divide between these two groups, and their participation costs are equal. Assume independent outcomes and use the variance formula for the difference in sample means.

With 18 participants in each group, the variance of the difference is $4^2/18+8^2/18=80/18\approx4.444$ outcome-units squared.

For a fixed $n_A+n_B$, minimizing $16/n_A+64/n_B$ gives $n_A/n_B=4/8=1/2$. Thus the allocation is $n_A=12$, $n_B=24$. Its variance is $16/12+64/24=4.000$ outcome-units squared, with standard error $\sqrt{4}=2.000$ outcome units. The balanced allocation’s standard error is $\sqrt{4.444}\approx2.108$ outcome units.

The unequal allocation improves precision for this specified difference because group B is more variable. The calculation does not establish that 12 and 24 is the best allocation for an omnibus test involving additional groups, or that the classical equal-variance F test is suitable when the variances really differ. In practice, variance estimates used for planning are uncertain, so a modest adjustment may be more sensible than treating their estimated ratio as exact.

Accounting for Different Participation Costs

When costs differ, the constraint may be a fixed budget rather than a fixed number of participants. For a specified difference between two independent groups, suppose the cost per participant is $c_i$ in group $i$ and $c_j$ in group $j$. With variances treated as planning values, minimizing $\sigma_i^2/n_i+\sigma_j^2/n_j$ subject to $c_i n_i+c_j n_j=B$ gives

n_i\ \propto\ \frac{\sigma_i}{\sqrt{c_i}}, \qquad \frac{n_i}{n_j} =\frac{\sigma_i}{\sigma_j}\sqrt{\frac{c_j}{c_i}}.

This rule balances the precision gained by adding a participant against that participant’s cost. It favors a group with greater variability, which needs more observations for precision, and a group with lower cost, which can be sampled more readily. As above, it targets the precision of one specified difference; a different objective can call for a different allocation.

Worked Example: Allocating a Fixed Budget

A hypothetical study aims to estimate the difference between groups A and B. Planning values are $\sigma_A=\sigma_B=5$ outcome units, participant costs are $c_A=\$20$ and $c_B=\$80$, and the budget is $\$1{,}800$. Assume independent outcomes and use the cost-constrained precision rule for this difference.

The recommended ratio is $n_A/n_B=(5/5)\sqrt{80/20}=2$. Combining $n_A=2n_B$ with the budget constraint gives $20(2n_B)+80n_B=1{,}800$, so $120n_B=1{,}800$, $n_B=15$, and $n_A=30$.

The variance of the estimated difference is $25/30+25/15=2.500$ outcome-units squared. For comparison, an allocation of 18 per group uses the same budget and has variance $25/18+25/18\approx2.778$ outcome-units squared. Under these planning values, assigning more participants to the less expensive group improves precision for the difference. Rounding or practical constraints may require a nearby feasible allocation.

Why There Is No Universal Omnibus Optimum

For the classical one-way test, the noncentrality parameter makes the limitation of a simple allocation rule explicit. It depends both on the sample sizes and on the actual pattern of population means. Equal allocation is often a sensible, symmetric choice when the effect pattern is unknown or when all groups should be treated comparably. But if a particular alternative is specified in advance, a different allocation can produce a larger $\lambda$, even when variances and costs are equal.

Worked Example: Power Depends on the Mean Pattern

Consider a hypothetical one-way ANOVA with three normally distributed groups, common variance $\sigma^2=1$ outcome-units squared, independent observations, and $N=30$. Suppose the planning alternative is $(\mu_1,\mu_2,\mu_3)=(0,1,4)$ outcome units. Compare balanced allocation $(10,10,10)$ with $(14,2,14)$, using the noncentrality formula.

For the balanced allocation, $\bar{\mu}_n=(10(0)+10(1)+10(4))/30=5/3$. Therefore, $\lambda=[10(0-5/3)^2+10(1-5/3)^2+10(4-5/3)^2]/1=86.667$.

For allocation $(14,2,14)$, the weighted mean is $\bar{\mu}_n=(14(0)+2(1)+14(4))/30=58/30=1.933$. Thus $\lambda=[14(0-1.933)^2+2(1-1.933)^2+14(4-1.933)^2]/1\approx113.867$.

For this particular alternative, the unbalanced allocation has the larger noncentrality parameter and therefore greater power at the same $\alpha$. The calculation assumes the stated normal, independent, common-variance model and treats the means as known planning values. It is not a general recommendation: giving group 2 only two participants would be unattractive if the mean pattern were uncertain or if that group were important in its own right.

Choosing an Allocation in Practice

Allocation should follow the study’s objective rather than a slogan about balance. Before finalizing it, identify what is fixed (a participant total, a budget, or both), whether group variances and costs are expected to differ, and whether any particular comparisons or groups deserve priority. Use plausible planning values, but recognize that anticipated means and variances are estimates rather than certainties.

Similar variances and costs, no special priority: use equal or nearly equal group sizes. This balances precision across group-mean differences.
Different variances: consider allocating more participants to groups with greater variability when precision for relevant differences is the objective; ensure the eventual analysis is appropriate for the variance pattern.
Different costs: account for the budget constraint. For a specified two-group difference, the cost-adjusted rule favors larger, less costly groups as well as more variable groups.
Specified mean pattern: allocation can change omnibus power through $\lambda$. A design tuned to one anticipated pattern may be less effective for another, so state the planning assumptions clearly.

Common Mistake

A common mistake is to say, “Equal allocation always maximizes power, so every group must have the same size.” The reasoning confuses a strong general-purpose design choice with a universal optimum. Equal allocation minimizes the total pairwise variance under equal variances and costs, but omnibus power also depends on the alternative mean pattern. Unequal variances or costs can change precision priorities, and a specified alternative can favor an unbalanced allocation. The correction is to state the objective and assumptions, then explain what the chosen allocation optimizes.

Key takeaway: With equal variances and equal costs, balanced groups provide symmetric, efficient precision across group comparisons. Unequal variances, unequal costs, and anticipated mean patterns can alter the best allocation. No allocation rule is optimal for every objective and every alternative.

Check Your Understanding

Use the model assumptions and allocation objective in each answer.

For $k$ groups and fixed $N$, why does equal allocation minimize the sum of pairwise sample-mean difference variances when all groups have the same variance?
Two groups have planning standard deviations of 3 and 6, equal participant costs, and a fixed combined sample size. Which group receives more participants under the precision rule for their difference, and in what ratio?
For a specified two-group difference, explain why the cost-adjusted allocation favors the less expensive group when variances are equal.
Does equal allocation necessarily maximize omnibus power for every specified pattern of group means? Explain using the noncentrality parameter.
Why should unequal group variances prompt attention to the analysis method as well as the allocation?