Two Different Beneficiaries of an Allocation Decision
An allocation target can be called “ethical” for more than one reason. It may direct more enrolled participants to the arm with higher expected utility. It may instead preserve the information needed to estimate a treatment contrast reliably, so that later patients and clinicians can make better decisions. These aims can point to different allocations. Distinguishing them matters because a design that benefits the average participant enrolled in the trial does not automatically produce the strongest evidence, and a design that improves precision does not necessarily maximize the welfare of participants in the trial.
The distinction is about whose benefit is being counted and when that benefit is expected. Patient-level benefit concerns the outcomes or utilities of people enrolled now. Trial-level benefit concerns the value of the evidence the trial produces, including its possible value to future patients. Evidence can influence later care only if it is reliable, relevant, and used; its value is therefore not a guaranteed clinical benefit to any particular future person.
A Simple Two-Arm Framework
Consider a trial with \(N\) participants and two arms: treatment and control. Let \(q\) be the fraction assigned to treatment, so the control fraction is \(1-q\), where \(0<q<1\). For the basic calculations below, assume the total sample size is fixed, outcomes from different participants are independent, and each arm’s outcome distribution is stable over the trial. Assume also that treatment and control assignment do not change the outcomes of other participants. These assumptions make the expected utility and precision calculations transparent; violations can change which allocation is appropriate.
Let \(u_1\) and \(u_0\) denote the expected utility for an enrolled participant assigned to treatment and control, respectively. Utility can incorporate outcome benefit, treatment-related harm, and costs, provided it is defined consistently across arms. The expected utility per enrolled participant under fraction \(q\) is the weighted average of arm-specific expected utilities.
This is a patient-level allocation objective: it asks how the allocation changes the expected utility of the people enrolled. If \(u_1>u_0\), \(W(q)\) increases with \(q\); if \(u_1<u_0\), it decreases. Without restrictions, the largest value occurs at an endpoint, assigning everyone to the arm with higher expected utility. But then the trial cannot estimate a between-arm contrast from concurrent randomized groups. A positive welfare slope alone does not settle whether such an allocation is acceptable or scientifically useful.
A common measure of trial-level evidence is the variance of the difference between the two sample means. Suppose the outcome is quantitative, with population variances \(\sigma_1^2\) under treatment and \(\sigma_0^2\) under control. If \(n_1=Nq\) participants receive treatment and \(n_0=N(1-q)\) receive control, the variance of the difference in sample means is
This formula assumes independent observations, finite arm-specific variances, and a comparison based on the two sample means. Smaller variance means greater precision for this contrast at the fixed sample size. It does not capture every dimension of trial value: for example, it does not by itself measure whether the chosen outcome is clinically important or whether the result will apply to future patients.
Why Precision Can Favor a Different Allocation
For a fixed \(N\), minimizing the variance is equivalent to minimizing \(g(q)=\sigma_1^2/q+\sigma_0^2/(1-q)\). Differentiating for \(0<q<1\) gives \(g'(q)=-\sigma_1^2/q^2+\sigma_0^2/(1-q)^2\). When \(\sigma_1^2>0\) and \(\sigma_0^2>0\), both standard deviations and denominators are positive, so taking positive square roots of \(\sigma_1^2/q^2=\sigma_0^2/(1-q)^2\) gives \(\sigma_1/q=\sigma_0/(1-q)\). Solving yields \(q=\sigma_1/(\sigma_1+\sigma_0)\). Further, \(g''(q)=2\sigma_1^2/q^3+2\sigma_0^2/(1-q)^3>0\), so this point is the unique minimum. If exactly one variance is zero, \(g(q)\) has no interior minimum and its infimum is approached at the boundary favoring the zero-variance arm; if both variances are zero, \(g(q)=0\) for every \(0<q<1\).
This precision calculation is the Neyman allocation for a two-arm comparison. It presumes that the objective is precision for this particular contrast and that the stated variance model is suitable. The earlier material on ethical targets and Neyman allocation introduces the contrast between these kinds of targets; here the central point is that precision serves the trial’s evidential objective, whereas \(W(q)\) serves the enrolled participants’ expected utility objective.
Worked Example: Patient Utility Favors More Treatment
Suppose, hypothetically, that a binary outcome is success or failure and that success has utility 1 while failure has utility 0 in either arm. Assume the success probabilities are \(p_1=0.70\) under treatment and \(p_0=0.55\) under control, with no additional harm or cost in this simplified utility measure. Compare a treatment fraction of \(q=0.50\) with \(q=0.80\) for \(N=100\).
The expected utility per participant is the expected success probability: \(W(q)=q(0.70)+(1-q)(0.55)=0.55+0.15q\). At \(q=0.50\), \(W(0.50)=0.55+0.15(0.50)=0.625\), or 62.5 expected successes per 100 participants. At \(q=0.80\), \(W(0.80)=0.55+0.15(0.80)=0.670\), or 67.0 expected successes per 100 participants.
Under these assumptions, assigning 80 rather than 50 participants to treatment raises the expected number of successes by \(67.0-62.5=4.5\). This is an expectation across hypothetical repetitions, not a promise that exactly 4.5 additional people will succeed in one trial. Nor does the calculation establish that \(q=0.80\) is the best trial design: it says nothing yet about the precision required to estimate the treatment-control difference.
Worked Example: A Precision Target Is Not a Welfare Target
Suppose, hypothetically, that a continuous outcome has treatment variance \(\sigma_1^2=0.21\) and control variance \(\sigma_0^2=0.25\). For a fixed total sample size, calculate the precision allocation.
The arm-specific standard deviations are \(\sigma_1=\sqrt{0.21}\approx0.4583\) and \(\sigma_0=\sqrt{0.25}=0.5000\). Therefore,
The variance-minimizing allocation is about 47.82% to treatment and 52.18% to control. For \(N=200\), this corresponds approximately to 95.64 treatment and 104.36 control participants before rounding to feasible integer counts. The result follows from the variances alone. It does not say that control is more effective or more beneficial: it says that, for these arm variances and this estimator, a slightly larger control group minimizes the variance of the difference in means.
To see the distinction, suppose the expected utilities are \(u_1=0.70\) and \(u_0=0.55\), as in a hypothetical success-utility scale. The welfare function increases with \(q\), while the precision target is \(q\approx0.4782\). The two targets answer different questions, so their difference is not a contradiction.
Combining the Aims With an Explicit Constraint
A design can prioritize participant utility while requiring enough precision for a useful comparison. This is not the same as claiming that one allocation simultaneously maximizes both objectives. It is a transparent compromise: specify a precision requirement, then maximize expected utility among allocations that meet it. The constraint must be justified by the trial’s scientific purpose and not chosen merely because it produces a preferred allocation.
Worked Example: Maximize Utility Subject to a Precision Requirement
Use the hypothetical variances \(\sigma_1^2=0.21\) and \(\sigma_0^2=0.25\), with \(N=200\), and require the standard error of the difference in means to be at most 0.0700. Suppose also that \(u_1=0.70\) and \(u_0=0.55\). Assume the independent-outcome variance model above applies and that fractional allocations are a planning approximation; actual arm counts must be integers.
The requirement \(SE\leq0.0700\) is equivalent to variance at most \(0.0700^2=0.004900\). Substituting the variance formula and multiplying by \(N=200\), the allowable fractions satisfy
For \(0<q<1\), multiplication by \(q(1-q)>0\) preserves the inequality. Rearranging gives \(0.21+0.04q\leq0.9800q(1-q)\), or \(0.9800q^2-0.9400q+0.2100\leq0\). The roots are
Because the quadratic opens upward, the constraint holds approximately for \(0.3542\leq q\leq0.6050\). The welfare function is \(W(q)=0.55+0.15q\), which increases with \(q\). Thus, among allocations meeting the precision requirement, the largest feasible treatment fraction maximizes expected utility. Taking \(q=0.6049\), close to the upper boundary but within rounding tolerance, gives approximately \(n_1=120.98\) and \(n_0=79.02\). The variance is \(0.21/120.98+0.25/79.02\approx0.004900\), so the standard error is approximately \(0.0700\). Expected utility per participant is \(0.55+0.15(0.6049)=0.6407\), corresponding to about \(200(0.6407)=128.1\) expected successes under this utility model.
This target differs from both the unconstrained welfare endpoint \(q=1\) and the precision-only target \(q\approx0.4782\). It is defensible only if the precision threshold is scientifically meaningful and the assumed variances and utilities are credible. Integer rounding and any minimum-arm requirements should be checked in the final design.
Diagnostics, Alternatives, and Limits
Before interpreting an allocation target, check what objective it actually optimizes. A useful diagnostic is to write the objective in words and in symbols: is it expected utility among enrolled participants, variance of a particular estimator, or a constrained combination? Then inspect the inputs. Welfare calculations are sensitive to the utility scale, response-rate assumptions, treatment harms, and costs. Precision calculations are sensitive to the outcome variance model, sample size, and the chosen estimand. If these inputs are uncertain, a point target can give a false impression of certainty; sensitivity analysis can show whether reasonable alternatives materially change the allocation.
The main alternatives can be summarized as follows:
| Allocation objective | What it prioritizes | What it does not guarantee |
|---|---|---|
| Patient-welfare target | Expected utility among enrolled participants | Precise estimation of the treatment contrast |
| Precision target | Information for a specified comparison | Highest expected utility for enrolled participants |
| Constrained or trade-off target | A stated balance, such as maximizing utility subject to adequate precision | That the chosen constraint or trade-off is uniquely correct |
Several limitations deserve emphasis. First, a trial’s broader evidential value is not fully represented by the standard error of one treatment contrast. Validity, outcome relevance, follow-up, generalizability, and the chance that findings inform care all matter. Second, the calculations here treat arm-specific outcomes and utilities as stable. Time trends, changing eligibility, interference between participants, or inaccurate response and variance estimates may undermine that simplification. Third, an allocation target is a design objective, not by itself a complete randomization procedure or a guarantee about the realized sample. Finally, ethical and scientific constraints can rule out allocations that look attractive under a single mathematical objective.
A frequent mistake is to say, “Treatment has higher expected utility, so assigning as many participants as possible to treatment is automatically the most ethical trial design.” That reasoning considers only the expected outcomes of current enrollees under the specified utility model. It ignores whether the trial can answer its question and whether reliable evidence may benefit future patients. The correction is to state the patient-level objective and the evidence requirement separately, then explain how the design reconciles them. The reverse mistake is to treat the precision-minimizing allocation as ethically optimal simply because it improves one standard error; precision alone does not measure participant welfare.
Check Your Understanding
Answer each question, then compare with the brief solutions below.
- If \(u_1=0.60\) and \(u_0=0.50\), what is the direction of change in \(W(q)\) as \(q\) increases? What does that answer not establish?
- For \(N=100\), \(\sigma_1=2\) and \(\sigma_0=1\), calculate the precision-minimizing treatment fraction.
- In one sentence, distinguish trial-level benefit from a direct benefit to an enrolled participant.
- Give one reason why a precision-only allocation could be a poor choice even if its variance calculation is correct.
Solutions. 1. Since \(W(q)=0.50+0.10q\), expected utility increases with \(q\); this does not establish that an all-treatment trial is scientifically adequate or ethically acceptable. 2. \(q_{\mathrm{prec}}=2/(2+1)=0.6667\), so about 66.67% is assigned to treatment. 3. Trial-level benefit is the potential value of reliable evidence for decisions, often including future care; a direct participant benefit is an expected outcome or utility for someone enrolled now. 4. For example, the target may ignore treatment harms, costs, or the utilities of enrolled participants, or may optimize precision for an inappropriate estimand.