From an Ethical Principle to a Usable Target
An ethical allocation target is not just a ratio that appears favorable. It is the endpoint of a decision: whose welfare counts, what outcomes contribute to that welfare, how those outcomes are valued, and what ethical and inferential requirements limit the allocation. A target can be calculated correctly yet remain hard to defend if these choices are hidden.
The purpose here is to make those choices explicit and auditable. Earlier material on patient benefit, minimum control allocation, precision constraints, and treatment-related harm provides the ingredients. The task now is to assemble them into a documented objective and check that the resulting number answers the intended question.
Specify the Objective and the Constraints
Begin by defining a participant-level utility \(U\) on a stated scale and time horizon. Utility may summarize relevant benefits, harms, and costs, but the components and their weights must be justified for the decision at hand. Let \(\mu_1=E(U\mid\text{treatment})\) and \(\mu_0=E(U\mid\text{control})\) be the expected utilities under assignment to each arm for the trial population. Under the assumption that the marginal utility of each enrolled participant is the quantity being maximized, the expected utility per participant at target \(q\) is
This is a participant-level welfare objective. It does not automatically include the value of learning from the trial, nor does it represent a separate objective such as maximizing precision. Those considerations must be expressed as constraints or included in a clearly justified objective.
Write each constraint in operational terms. Ethical requirements might include a minimum fraction for each arm, a limit on how many participants may receive an arm with known disadvantage, or a requirement that the target remain within a prespecified range. Inferential requirements might specify a maximum planning standard error for the treatment-control contrast, a minimum information level per arm, or a sample size and power criterion. State the primary outcome or utility measure used for inference: a precision constraint for one outcome does not guarantee adequate precision for another.
The calculation also needs assumptions. For the simple two-arm planning calculation below, assume a fixed total sample size \(N\), independent participants, a common target population, a well-defined treatment and control, and no interference between participants. Assume the planned outcome variance in each arm is represented by \(\sigma_1^2\) and \(\sigma_0^2\). These may be planning values rather than known population quantities. The variance formula is a planning approximation under the proposed allocation; the actual adaptive design and estimator must still be assessed on their own terms.
Worked Example: Maximizing Utility Subject to a Precision Requirement
Worked Example: Maximizing Utility Subject to a Precision Requirement
Hypothetical planning scenario. A trial plans to enroll \(N=200\) participants. On a prespecified utility scale, the expected treatment utility is \(\mu_1=0.70\) points and the expected control utility is \(\mu_0=0.50\) points. The planning variances are \(\sigma_1^2=0.21\) and \(\sigma_0^2=0.25\), in squared utility points. At least 20% of participants must be assigned to each arm, and the planning standard error for the difference in mean utility must not exceed \(0.08\) utility points. Find the welfare-maximizing feasible target.
Check the objective. The expected utility is \(W(q)=0.70q+0.50(1-q)=0.50+0.20q\). Because its slope is \(0.20>0\), expected utility increases as \(q\) increases. Thus, among feasible targets, choose the largest \(q\).
Apply the minimum-arm requirement. At least 20% in each arm gives \(q\geq0.20\) and \(1-q\geq0.20\), or \(0.20\leq q\leq0.80\).
Apply the precision requirement. Under the stated independent-arm planning model, the variance of the difference in sample means is \(0.21/(200q)+0.25/[200(1-q)]\). Requiring its square root to be at most \(0.08\) gives
For \(0<q<1\), multiplying by the positive quantity \(q(1-q)\) preserves the inequality: \(0.21(1-q)+0.25q\leq1.28q(1-q)\). Rearranging gives
The roots are \([1.24\pm\sqrt{1.24^2-4(1.28)(0.21)}]/(2.56)=[1.24\pm0.68]/2.56\), namely \(0.21875\) and \(0.75\). Since the quadratic opens upward, the precision constraint holds for \(0.21875\leq q\leq0.75\). Intersecting this with the ethical range gives the same interval. The welfare objective is increasing, so \(q^*=0.75\).
At \(q=0.75\), the planned counts are 150 treatment and 50 control participants. The planning standard error is \(\sqrt{0.21/150+0.25/50}=\sqrt{0.0064}=0.08\) utility points, exactly the allowed maximum. The target therefore meets both stated requirements; it does not establish that the assumed utilities or variances are correct.
Why the Utility Definition Can Change the Decision
A target is conditional on its objective, not an ethical verdict independent of that objective. In particular, including harm or cost can change which arm has higher expected utility. The next example uses the same feasible interval derived above to show how documenting the utility definition affects the result.
Worked Example: Rechecking the Target After Including Burden
Hypothetical scenario. Suppose a committee initially compares a benefit score alone and records expected scores of 8 points for treatment and 5 points for control. On that measure, the treatment arm has the higher mean. Before setting the target, the committee decides that treatment-related burden and cost must also be reflected: the treatment mean has a burden of 2.5 points and a cost of 1 point, while the control mean on the resulting net-utility scale is 5.5 points. Treat these valuations as planning inputs, not as established facts.
The treatment mean net utility is \(8-2.5-1=4.5\) points, compared with \(\mu_0=5.5\) points for control. Thus \(W(q)=4.5q+5.5(1-q)=5.5-q\), which decreases as \(q\) increases. If the ethical and precision constraints are those in the preceding example, feasible values remain \(0.21875\leq q\leq0.75\). The welfare-maximizing target is now the smallest feasible value, \(q^*=0.21875\), rather than the largest.
At this target, the expected net utility is \(5.5-0.21875=5.28125\) points per participant. The implied treatment fraction is about 21.9%, with about 78.1% allocated to control before rounding. The target still satisfies the stated precision threshold: substituting \(q=0.21875\) into the planning variance gives a standard error of \(0.08\) points. This calculation illustrates why a report must disclose the utility components and weights. A benefit-only score and a net-utility score answer different decision questions.
Derivation, Feasibility, and Reporting
The core decision follows directly from the objective. For the linear participant-level welfare function,
If \(\mu_1-\mu_0>0\), welfare increases with \(q\), so the constrained optimum is the largest feasible \(q\). If \(\mu_1-\mu_0<0\), it is the smallest feasible \(q\). If the difference is zero, all feasible targets give the same value of this particular welfare objective; another stated criterion must distinguish them. This conclusion depends on the linear objective and fixed utility values. It need not hold for nonlinear welfare functions or when the utilities themselves depend on allocation.
A feasible set may be empty. For example, a stringent precision threshold can conflict with minimum-arm requirements or the planned sample size. In that case, do not report a target as though both requirements were met. Revisit the sample size, threshold, or ethical restrictions transparently, and state which requirements are retained. Likewise, if utility inputs are uncertain, report sensitivity analyses across plausible values rather than giving a single target an unwarranted appearance of certainty.
For the variance-based precision constraint, check the inputs and the boundary, not only the final number. The formula requires \(0<q<1\); as either arm fraction approaches zero, the corresponding variance contribution grows without bound when that arm has positive variance. Check that both arm variances are on the scale of the outcome used for inference, that \(N\) and the planned arm counts agree, and that rounding does not violate a minimum fraction or precision requirement.
Worked Example: Checking a Rounded Target Against the Planned Sample Size
Hypothetical planning check. Suppose the same utility variances and maximum standard error of \(0.08\) points are proposed, but the planned total sample size is \(N=120\). A proposed target of \(q=0.75\) would assign 90 participants to treatment and 30 to control, satisfying the 20% minimum in each arm. Its planning standard error, however, is \(\sqrt{0.21/90+0.25/30}=\sqrt{0.010667}=0.1033\) utility points, which exceeds \(0.08\).
Thus, \(q=0.75\) is not feasible under the proposed precision requirement for this sample size, even though it was feasible at \(N=200\). At \(q=0.75\), the variance is \(1.28/120=0.010667\), so the sample size needed for a standard error no greater than \(0.08\) at that allocation is at least \(1.28/0.08^2=200\). The committee would need a different feasible target, a larger sample, or a clearly justified revision of the inferential requirement. Simply rounding the target cannot resolve the conflict.
A useful report distinguishes the target from the mechanism that may later implement it. State the intended long-run or planned fraction, not an unqualified probability for the next assignment. The operational randomization procedure, its adaptation, and its inferential properties require their own specification; the target alone does not determine them.
Comparison with alternatives clarifies what the selected number does and does not optimize. Equal allocation is a simple reference point and may support precision, but it does not necessarily maximize the stated patient-utility objective. Unconstrained welfare maximization can drive allocation toward an endpoint, which may conflict with control-group ethics or information needs. A constrained ethical target makes those restrictions explicit. A variance-oriented allocation, such as Neyman allocation, focuses on precision for a specified contrast; it need not maximize participant utility. None is universally preferable: the choice depends on the declared objective and the constraints that must be respected.
Common Mistakes and Practical Diagnostics
- Reporting a ratio without its objective. A number such as 3:1 is not interpretable as an ethical decision unless the utility scale, population, and constraints are given. Report the treatment and control fractions and the rationale.
- Calling a precision constraint an ethical objective. Precision limits what targets are allowed; it does not say whose welfare is being maximized. Name the two roles separately.
- Mixing outcome scales. Do not subtract a cost measured in dollars from a clinical score without specifying a conversion or utility valuation that puts them on a common scale.
- Treating planning inputs as known truths. Assumed means and variances can be uncertain or misspecified. Show their source or rationale and examine how plausible changes affect the target.
- Ignoring implementation and rounding. A continuous target may not yield exact integer counts. Verify the rounded allocation against minimum-arm and precision constraints, and describe any planned rounding convention.
- Interpreting the target as a guarantee. A target does not guarantee realized proportions, patient benefit, or inferential performance. Those depend on the design, implementation, assumptions, and observed data.
Check Your Understanding
Answer each question, then compare your reasoning with the solutions.
- If \(W(q)=0.40+0.12q\), and the feasible interval is \(0.25\leq q\leq0.70\), what target maximizes this welfare objective? State why.
- A report gives a target of 70% treatment and 30% control but does not state the utility scale or precision requirement. Name two pieces of information needed to interpret or evaluate the target.
- For fixed \(N=100\), planned variances \(\sigma_1^2=0.16\) and \(\sigma_0^2=0.25\), and \(q=0.60\), calculate the planning standard error of the difference in mean outcomes.
- Why is a target that satisfies a planning standard-error constraint not, by itself, evidence that the randomization procedure will achieve the desired inferential performance?
Solutions. (1) \(W(q)\) has positive slope \(0.12\), so it increases over the feasible interval; the maximizing target is \(q=0.70\). (2) For example, the report should specify how participant utility is defined and valued, and what inferential criterion or precision threshold restricts the target. It should also disclose assumptions and the minimum-arm requirements. (3) The variance is \(0.16/(100(0.60))+0.25/(100(0.40))=0.002667+0.00625=0.008917\). Therefore, \(SE=\sqrt{0.008917}=0.09443\) outcome units. (4) The calculation is based on planning assumptions and a variance approximation under the target allocation. The implemented design, estimator, realized allocation, and actual outcome distributions may differ; these must be evaluated separately.