Ethical Allocation With a Minimum Control Group

Why Impose a Minimum Control Group?

If one arm is expected to provide greater patient utility than another, assigning more patients to that arm can increase the trial’s expected total utility. But a trial may also need a control group of a prespecified minimum size. That requirement rules out allocations with too few control participants, even if they would otherwise yield greater expected utility. The resulting problem is a simple constrained optimization: maximize expected utility over the allocations that satisfy the minimum.

The calculation is useful precisely because it makes the trade-off explicit. When treatment is more beneficial on average, every additional control participant has an opportunity cost in expected utility. A minimum control size specifies how much of that cost the design is willing to incur in exchange for retaining the required control group. The calculation does not, by itself, establish that the minimum is sufficient for valid inference or that expected utility captures every ethical concern.

Model and Feasible Allocations

Consider a two-arm trial with a fixed total of \(N\) participants. Let \(n_C\) be the number assigned to control and \(n_T\) the number assigned to treatment, so \(n_C+n_T=N\). Let \(u_C\) and \(u_T\) denote the known expected utility for a participant assigned to control and treatment, respectively. Utilities are measured in the same units and on a scale where larger values are preferred. The minimum control-group size \(m\) is a prespecified integer.

Assume \(N\) and \(m\) are fixed, with \(0\le m\le N\); each participant’s expected utility depends on assigned arm and not on the number or identities of other participants; and the arm-specific expected utilities are known and constant for this calculation. Additivity of expected utility means the expected total is the sum of the participant-level expected utilities. No distributional assumption about individual outcomes is needed to maximize this linear expected total.

Since \(n_T=N-n_C\), the objective can be written using control sample size alone:

Thus the slope with respect to \(n_C\) is \(u_C-u_T\). If that difference is negative, increasing the control group lowers expected total utility; if it is positive, increasing the control group raises utility; and if it is zero, the allocation does not change utility. Because the objective is linear and the feasible values form an integer interval, the maximum occurs at an endpoint unless all feasible allocations tie.

Deriving the Optimal Target

The one-participant change also proves the result directly. For any feasible \(n_C<N\), compare the allocation with \(n_C+1\) control participants:

When \(u_C-u_T<0\), every step to the right reduces utility, so the smallest feasible control count, \(m\), is optimal. When \(u_C-u_T>0\), every step to the right increases utility, so the largest feasible control count, \(N\), is optimal. When the difference is zero, every step leaves utility unchanged. This establishes the solution over the integer allocations, not just over a continuous approximation.

Worked Examples

Interpreting the Constraint and Checking the Calculation

The minimum \(m\) is a lower bound on the control sample size, not a requirement that the control group contain exactly \(m\) participants. It becomes the selected target when treatment has strictly higher expected utility because every additional control assignment then lowers the objective. If control has higher expected utility, the minimum does not bind at the optimum: the model instead assigns everyone to control. If utilities tie, the minimum limits the feasible set but does not determine a unique utility-maximizing allocation.

A practical check is to compute the change from transferring one participant between arms. Moving one participant from treatment to control must change expected total utility by \(u_C-u_T\). The sign should agree with the direction of the proposed optimum. A second check is to substitute the allocation counts into \(n_C+n_T=N\) and \(n_C\ge m\), then recompute the total utility from its two arm-specific contributions.

The minimum can be binding without guaranteeing that the control group is large enough for a desired precision or analysis. This objective contains no sampling-variance term, information requirement, or inferential constraint. The earlier material on ethical allocation as expected patient utility motivates the welfare objective; the earlier discussion of why pure welfare maximization can eliminate a control arm explains why a control requirement may be imposed. The present calculation answers the narrower question: given a specified minimum, which feasible allocation maximizes this expected total?

Comparison, Limitations, and Common Pitfall

The answer depends on the input utilities and on the model’s simplifying assumptions. If the values \(u_C\) and \(u_T\) are estimates rather than known quantities, uncertainty in them can change which arm appears preferable. If expected utility depends on patient characteristics, treatment effects vary across participants, or participants’ outcomes are dependent, the two fixed arm-level averages may not capture the allocation’s full consequences. In those settings, the simple linear objective may need to be revised.

Common mistake: “Treatment has higher utility, so the best allocation is to put everyone in treatment; the control minimum is only a suggestion.” This is wrong because the minimum is a feasibility condition, not a preference that can be ignored. When \(u_T>u_C\), the unconstrained utility maximum would place everyone in treatment, but that allocation violates \(n_C\ge m\) whenever \(m>0\). The constrained maximum instead uses exactly \(m\) control participants and \(N-m\) treatment participants.

Another mistake is to assume that any minimum control group automatically resolves the ethical or statistical design problem. It does not. It encodes one required feature of the design, while the utility calculation optimizes only the specified expected total. The result should be reported as a target under stated assumptions, not as a complete judgment about trial design.