Equal Allocation Across Several Experimental Arms

Why Equal Allocation Needs a Multi-Arm Calculation

A trial with several experimental treatments can use one common control group. Under equal allocation, the control and every treatment arm have the same number of participants. This is easy to describe and implement, but the recruitment total is not found by adding the sample sizes of separate two-arm trials: participants in the common control group are counted once, not once for every comparison.

The practical task is to turn the minimum sample-size requirements for each planned treatment-versus-control comparison into one common group size. If the comparisons have different requirements, the common size must meet the most demanding one. Equal allocation may therefore put some arms above their individual minimums; that is a consequence of requiring all arms to be the same size.

Assumptions and the Calculation Rule

This rule concerns a fixed, parallel-group trial with \(k\) treatment arms and one shared control arm. Participants are assigned to only one arm, and the same control participants are used in each treatment-versus-control comparison. The required sample sizes must come from planning calculations appropriate to the outcome, target effect, variability, significance level, power, and analysis being planned. Those inputs are treated as given here; equal allocation does not itself determine them.

The rule also assumes that the comparison-specific requirements refer to the same recruitment target—for example, the number of evaluable participants needed in each arm—and that the trial can recruit to a common group size. If each comparison has separate minimums for its control and treatment groups, all of those minimums must be met. The equal size must then be at least as large as the largest group-specific requirement, not merely the largest treatment requirement.

Here \(n_{C,\min}\) is the largest required control-group size among the comparisons, and \(n_{j,\min}\) is the required size for treatment arm \(j\). If each comparison instead supplies one common per-arm minimum \(m_j\), use \(n_{\mathrm{equal}}=\max_j m_j\). When a calculation gives a fractional minimum, round that minimum up to an integer before finding the maximum. Rounding down could leave an arm below its requirement.

Why the Maximum Is the Smallest Feasible Equal Size

The result follows directly from the constraints. Equal allocation requires the same integer \(n\) in every arm. Meeting every group-specific minimum requires \(n\geq n_{C,\min}\) and \(n\geq n_{j,\min}\) for every treatment \(j\). Thus \(n\) must be at least the largest of those minimums. Choosing that largest minimum meets all the constraints, and choosing any smaller value would fail at least one. Multiplying by the number of arms gives the total.

This is a feasibility calculation, not an averaging calculation: an average of the minimums does not ensure that the most demanding comparison is adequately sized. Nor should the control be counted separately for every treatment when calculating total recruitment. There is one common control group in the protocol.

When Control and Treatment Minimums Differ

Not every planning summary gives one per-arm minimum for each comparison. A comparison may instead specify different minimums for the control and its treatment—for example, because its design or analysis calls for different group sizes. Under equal allocation across the whole trial, the group size must meet every stated requirement, including every control requirement. Taking the largest number across the control and treatment minima is a simple way to enforce all of them.

Recruitment Targets and Allowance for Non-Evaluable Participants

A planning requirement may refer to participants whose outcomes can be included in the primary analysis rather than everyone randomized. If a proportion is expected to be non-evaluable, recruitment must be increased to allow for that loss. Apply the allowance to the required number per arm first, round up to a whole participant, and then multiply by the number of arms. The same allowance should not be added again to the total after the per-arm targets have already been inflated.

For a common anticipated non-evaluable proportion \(r\), a basic planning calculation is \(n_{\mathrm{recruit}}=\lceil n_{\mathrm{evaluable}}/(1-r)\rceil\), where \(\lceil\cdot\rceil\) means round up to the next integer. This is an allowance based on an anticipated rate, not a guarantee that the required number will be evaluable in every arm.

Checks, Alternatives, and Limitations

Before accepting a total, check that the number of arms is correct, the control is counted once, and every comparison-specific minimum is met. A useful audit is to write the proposed allocation explicitly and compare each arm with its requirement. Also check whether requirements mean randomized or evaluable participants, whether rounding has been applied upward, and whether all comparisons really use the same control group and recruitment population.

A tempting alternative is to calculate each two-arm comparison separately and add the resulting totals. That would count the shared control repeatedly, so it does not give the recruitment total for one multi-arm trial. Another tempting approach is to average the per-comparison minimums and use that as the common size. This can leave the arm associated with the largest requirement under its minimum. Equal allocation is straightforward, but it may recruit more participants than a plan that allows different group sizes. Whether a different allocation is preferable depends on the design objectives and constraints; it is not established merely by this equal-allocation calculation.

Equal allocation also has limits. It does not establish that the sample-size requirements themselves are appropriate, account automatically for multiple comparisons, or resolve whether sharing a control is suitable for the scientific question. If treatment groups differ in outcome variability, follow-up, or anticipated missingness, one common recruitment target may need further justification. And if the shared control data are not genuinely common to the comparisons—for example, because eligibility or assessment differs—simply counting one control group may not be defensible.