Allocation Ratios and Sample-Size Requirements

Why Allocation Ratios Matter

In a multi-arm trial, each treatment-versus-control contrast uses observations from two groups. The number assigned to each group therefore affects the contrast's precision. In a MAMS design, this calculation applies to the cumulative sample available for the relevant comparison at a given analysis; the allocation ratio alone does not determine how many participants will actually be enrolled if the design can stop early.

With several treatment arms and one shared control, increasing the control group can improve every treatment-control comparison. But those gains must be weighed against the participants assigned to the treatment arms. The central calculation is the variance of a difference in sample means, followed by solving that variance for the enrollment needed to attain a specified precision.

Definition: Let \(n_0\) be the number of participants in the control arm and \(n_i\) the number in active arm \(i\), for \(i=1,\ldots,m\). The allocation ratio for arm \(i\), relative to control, is \(r_i=n_i/n_0\). Thus \(n_i=r_i n_0\), and the total enrollment is \(N=n_0+\sum_{i=1}^m n_i=n_0(1+\sum_{i=1}^m r_i)\). If all active arms use the same ratio \(r\), then \(N=n_0(1+mr)\).

Contrast Variance With a Shared Control

Suppose the outcome for a participant in arm \(i\) has mean \(\mu_i\), and the control outcome has mean \(\mu_0\). Assume independent participants across and within arms, a common outcome variance \(\sigma^2\) in every arm, and complete outcome data. The estimated treatment-control contrast is \(\bar{x}_i-\bar{x}_0\), with target \(\mu_i-\mu_0\). Under these assumptions, the variance of a sample mean in arm \(i\) is \(\sigma^2/n_i\), and the two arm means are independent.

\operatorname{Var}(\bar{x}_i-\bar{x}_0) =\operatorname{Var}(\bar{x}_i)+\operatorname{Var}(\bar{x}_0) =\sigma^2\left(\frac{1}{n_i}+\frac{1}{n_0}\right).

The standard error is therefore \(\sigma\sqrt{1/n_i+1/n_0}\). Substituting \(n_i=r_i n_0\) makes the allocation's role explicit: the variance is \(\sigma^2(1+1/r_i)/n_0\). A smaller treatment-to-control ratio means fewer observations in that treatment arm and a larger variance for its contrast, if the control size is held fixed.

The estimates for different treatment arms are not independent because they share the same control mean. For distinct active arms \(i\) and \(j\), independent active-arm means give:

\operatorname{Cov}(\bar{x}_i-\bar{x}_0,\bar{x}_j-\bar{x}_0) =\operatorname{Var}(\bar{x}_0) =\frac{\sigma^2}{n_0}.

This positive covariance matters for joint inference across arms. It does not change the variance of any one pairwise contrast, but it does mean that a procedure for simultaneous inference should account for the shared-control dependence. If a calibrated simultaneous critical value \(c\) is being used, the planned half-width for an individual contrast is \(c\,SE\); for an ordinary two-sided confidence interval based on a normal approximation, \(c=z_{1-\alpha/2}\).

Worked Example: Contrast Precision and Shared-Control Covariance

Consider a hypothetical three-arm comparison with one control arm and two active arms. Suppose \(n_0=60\), \(n_1=90\), and \(n_2=45\), and assume a common outcome standard deviation of \(\sigma=15\) units. Check that the arm sizes and variance assumptions are the inputs required for the calculation.

For active arm 1, the estimated contrast variance and standard error are:

\widehat{\operatorname{Var}}(\bar{x}_1-\bar{x}_0) =15^2\left(\frac{1}{90}+\frac{1}{60}\right) =225(0.0277778)=6.250, \qquad SE_1=\sqrt{6.250}=2.500\text{ units}.

For active arm 2:

\widehat{\operatorname{Var}}(\bar{x}_2-\bar{x}_0) =225\left(\frac{1}{45}+\frac{1}{60}\right) =225(0.0388889)=8.750, \qquad SE_2=\sqrt{8.750}=2.958\text{ units}.

Arm 2 has the larger standard error because it has fewer participants than arm 1 while using the same control group. The covariance between the two estimated contrasts is \(225/60=3.750\) squared units. Their correlation is \(3.750/(2.500\times2.958)=0.507\). Thus the comparisons are positively correlated, not independent, even though participants in the two active arms are separate.

Turning a Precision Target Into Enrollment

A precision target can be stated as a desired confidence-interval half-width \(h\), in the outcome's units. For a two-sided normal-approximation interval with specified critical value \(c\), require \(c\,SE_i\le h\). With a common allocation ratio \(r\), each contrast has the same planned variance if the active-arm sizes are equal. Substituting the ratio-based variance gives a direct requirement for the control size and total enrollment.

c\,\sigma\sqrt{\frac{1+1/r}{n_0}}\le h \quad\Longrightarrow\quad n_0\ge \frac{c^2\sigma^2(1+1/r)}{h^2}, \qquad N=n_0(1+mr).

For unequal ratios \(r_i\), apply the same condition to each contrast separately. To meet the target for every arm, choose \(n_0\) at least as large as the largest arm-specific requirement, \(c^2\sigma^2(1+1/r_i)/h^2\), then calculate each \(n_i=r_i n_0\). The arm with the smallest ratio is the limiting contrast under the common-variance assumptions.

Worked Example: Enrollment for a Target Half-Width

Suppose a hypothetical trial has three active arms, each allocated equally to control (\(m=3\), \(r=1\)). The assumed common outcome standard deviation is 12 units, and the desired two-sided 95% normal-approximation confidence-interval half-width for each contrast is at most 4 units. Use \(c=1.96\). Assume the variance estimate is suitable for planning and the normal approximation is appropriate.

The required control size is:

n_0\ge \frac{1.96^2(12^2)(1+1)}{4^2} =\frac{3.8416(144)(2)}{16} =69.1488.

The continuous calculation gives \(n_0=69.1488\) and \(n_i=69.1488\) in each active arm, for \(N=4(69.1488)=276.5952\). Because participant counts must be integers, rounding the required group size upward gives 70 participants in each arm, or \(N=280\). The resulting standard error is \(12\sqrt{1/70+1/70}=2.028\) units, and the half-width is \(1.96(2.028)=3.976\) units. This is below the 4-unit target. Rounding the continuous requirement downward would not be justified by the inequality.

Choosing Ratios When Total Enrollment Is Fixed

A ratio can be selected in advance for practical or scientific reasons, but it can also be compared in terms of contrast precision at fixed total enrollment. For one treatment-control contrast with equal variances and a fixed two-arm total, equal allocation minimizes its variance. With multiple equally important active arms sharing one control, however, the best common ratio for minimizing the common contrast variance is generally not 1:1.

To see the allocation trade-off, let each of \(m\) active arms have size \(x\), let the control size be \(y\), and hold total enrollment \(N=y+mx\) fixed. The contrast variance apart from the common factor \(\sigma^2\) is \(1/x+1/y\). Substituting \(x=(N-y)/m\) gives \(m/(N-y)+1/y\). Differentiating with respect to \(y\) and setting the derivative to zero yields \(m/(N-y)^2=1/y^2\). The positive solution is \(N-y=\sqrt{m}\,y\), so the minimizing sizes and ratio are:

y=\frac{N}{1+\sqrt{m}}, \qquad x=\frac{N}{\sqrt{m}(1+\sqrt{m})}, \qquad r=\frac{x}{y}=\frac{1}{\sqrt{m}}.

This result assumes equal importance for the active-arm contrasts, a common variance, and a fixed total sample size. It is not a universal allocation rule: unequal treatment priorities, different variances, costs, or other design objectives can change the choice. In a MAMS setting, the ratio also needs to be compatible with the planned stage sizes and the intended comparisons.

Worked Example: Comparing Ratios at 500 Participants

Suppose a hypothetical design has four active arms and one shared control, a common outcome standard deviation of 10 units, and a total enrollment of 500. Compare equal allocation across all five groups with a common treatment-to-control ratio of \(r=0.5\). Assume complete outcomes and a common variance. For both allocations, calculate the ordinary 95% normal-approximation half-width using \(c=1.96\).

With equal group sizes, there are 100 participants in each arm. A treatment-control contrast has standard error \(10\sqrt{1/100+1/100}=1.414\) units and half-width \(1.96(1.414)=2.772\) units.

For \(r=0.5\), the continuous allocation formula gives \(n_0=500/(1+4(0.5))=166.667\), with \(83.333\) participants per active arm. A feasible integer allocation close to this ratio is 168 control participants and 83 in each active arm: \(168+4(83)=500\). Its treatment-to-control ratio is approximately \(83/168=0.494\). The contrast standard error and half-width are:

SE=10\sqrt{\frac{1}{83}+\frac{1}{168}} =10\sqrt{0.0180006}=1.342\text{ units}, \qquad 1.96SE=2.630\text{ units}.

Under these assumptions, the unequal allocation gives a smaller individual contrast half-width than equal allocation at the same total enrollment. For four equally weighted active contrasts, the variance-minimizing common ratio from the derivation is \(1/\sqrt{4}=0.5\), which explains this result. It does not mean that the unequal allocation is preferable for every trial objective or every inference procedure.

Assumptions, Diagnostics, and Limitations

The calculations rely on a correctly specified outcome variance, independent participants, and the stated group sizes being the numbers with analyzable outcomes. A normal-theory calculation is exact for independent normal outcomes with known variance; with estimated variance or non-normal outcomes, it is generally a planning approximation whose adequacy depends on the design and sample sizes. If variance differs by arm, replace the common-variance expression with the relevant sum of arm-specific variances, such as \(\sigma_i^2/n_i+\sigma_0^2/n_0\) for contrast \(i\). The simple common-ratio result then need not optimize precision.

Before relying on a planned allocation calculation, check the following:

Group-size arithmetic: confirm that \(N=n_0+\sum_i n_i\), and that each stated ratio agrees with the arm counts.
Precision for every arm: calculate each \(c\,SE_i\), not just the contrast with the largest treatment group. The smallest active arm often determines the requirement.
Analysis critical value: use a critical value appropriate to the planned interval or inference procedure. A simultaneous critical value can be larger than 1.96, increasing the sample size needed for the same half-width.
Effective information: if attrition or unusable outcomes are anticipated, randomized group sizes may need to exceed the analyzable sizes used in the variance formula.
Dependence: account for shared-control covariance in joint inference; treating the treatment-control estimates as independent misrepresents their joint distribution.

A precision calculation is not automatically a power calculation. A confidence-interval half-width target controls the planned uncertainty under the assumed variance and sample sizes; power additionally depends on the effect being detected and the rejection rule. Also, total enrollment under a fixed allocation plan is different from expected enrollment when arms may stop early. Those quantities require continuation and stopping information beyond the allocation-ratio formulas developed here.

Common Mistakes

Using the ratio as the fraction of total enrollment. If \(r=n_i/n_0\), then with \(m\) equal active arms, \(n_0=N/(1+mr)\), not \(N/(1+r)\). The control group is shared, so all active-arm sizes enter the total.
Ignoring the shared control in a contrast variance. The variance includes both \(1/n_i\) and \(1/n_0\). More control participants improve each individual contrast, though comparisons using that control are correlated.
Rounding down a minimum group size. A continuous requirement must be rounded so the precision inequality still holds; recompute the standard error after integer allocation.
Assuming the ratio that is best for one contrast is automatically best for many. With multiple active arms sharing control, the fixed-total optimum under equal importance and common variance is \(r=1/\sqrt m\), not necessarily 1.

Key takeaway: For a common-variance treatment-control contrast, \(SE=\sigma\sqrt{1/n_i+1/n_0}\). Allocation ratios convert this into a direct enrollment requirement, while shared-control covariance must be recognized when the contrasts are considered jointly.

Check Your Understanding

Try each question, then check the concise solutions below.

There are two active arms with \(n_0=80\), \(n_1=80\), and \(n_2=40\). If the common standard deviation is 8 units, which contrast has the larger standard error, and what are the two standard errors?
With three active arms each allocated at \(r=0.5\), express total enrollment \(N\) in terms of \(n_0\).
For \(m=9\) equally important active contrasts, what common ratio minimizes their variance at fixed total enrollment under the equal-variance assumptions?
Why are two treatment-control contrast estimates positively correlated when their treatment arms are independent?

Solutions. (1) Arm 1: \(8\sqrt{1/80+1/80}=1.265\) units. Arm 2: \(8\sqrt{1/40+1/80}=1.549\) units; arm 2 has the larger standard error. (2) \(N=n_0(1+3(0.5))=2.5n_0\). (3) \(r=1/\sqrt{9}=1/3\). (4) Both contrasts subtract the same control mean, so their covariance is \(\operatorname{Var}(\bar{x}_0)=\sigma^2/n_0>0\).