Ethical Allocation as Expected Patient Utility

From Patient Outcomes to an Allocation Objective

An allocation target answers a design question: among the allocations under consideration, which one has the greatest expected value for the participants assigned in the trial? For an ethical allocation objective, that value is patient utility. The objective must specify both how likely each outcome is under each arm and how valuable each outcome is judged to be. Neither probabilities nor utility scores alone determine expected patient utility.

This is a design criterion, not a claim that an allocation will improve every participant’s outcome. Its result depends on the outcome probabilities used, the utility scale chosen, the population to which those probabilities apply, and which allocations are feasible. Utility scores encode value judgments as well as measurement choices; they should be stated clearly rather than treated as objective properties of outcomes.

The corresponding allocation shares are \(q_j=n_j/N\), with \(\sum_jq_j=1\). For a fixed total sample size \(N\), expected utility per assigned participant is $$ \frac{W(\mathbf n)}{N}=\sum_{j=1}^{J}q_j\mu_{U,j}. $$ Thus, the share assigned to an arm weights that arm’s expected utility in the trial-wide average. The total and per-participant objectives have the same maximizing allocations when \(N\) is fixed, because one is the other multiplied by \(N\).

Assumptions Behind the Objective

The calculation is straightforward, but its interpretation requires care. The outcome states must cover the outcomes included in the objective, and the probabilities for each arm must be nonnegative and sum to one. Each \(u_{jk}\) must represent the intended value of outcome \(k\) under arm \(j\), on a scale where the stated additions and probability-weighted averages are meaningful.

The probabilities should describe the population and outcome period relevant to the design. In a randomized trial, they can be interpreted as the probabilities of outcomes under assignment to each arm for eligible participants. Using them to compare allocations assumes the expected outcome under an arm does not change merely because a different number of participants are assigned to that arm. This is appropriate, for example, when one participant’s assignment does not affect another participant’s outcome. If there are important spillovers, or if treatment delivery changes with allocation, the arm-specific probabilities may need to depend on the allocation itself.

Utilities should also refer to a consistent perspective and time horizon. If one arm’s scores describe short-term outcomes while another’s represent long-term outcomes, their weighted averages do not answer a coherent allocation question. The model can allow \(u_{jk}\) to vary by arm, but any such variation needs a stated interpretation. It is not safe to assume that a higher probability of one named outcome automatically means higher expected utility.

Why the Formula Gives Expected Total Utility

For a participant assigned to arm \(j\), define the utility \(U\) to be \(u_{jk}\) if their outcome is state \(k\). The expected value of this participant’s utility is the sum of each possible utility multiplied by its probability: $$ E(U\mid j)=\sum_{k=1}^{K}P(Y(j)=k)u_{jk} =\sum_{k=1}^{K}p_{jk}u_{jk} =\mu_{U,j}. $$ If there are \(n_j\) participants assigned to arm \(j\), the expected sum of their utilities is \(n_j\mu_{U,j}\), by linearity of expectation. Summing this quantity over arms gives \(W(\mathbf n)\). This argument does not require the participants’ outcomes to be independent: linearity of expectation applies to the sum even when outcomes are dependent. Independence and other modeling assumptions may matter for assessing uncertainty, but they are not needed for this expected-value identity.

The derivation also shows what the objective leaves out: it adds expected utilities across participants. It does not, by itself, reward an allocation for estimating treatment effects precisely or for balancing arm sizes. Those are different design considerations.

Utilities Can Change the Ranking of Outcomes

An allocation objective is not determined by a success probability alone. An outcome’s contribution is its probability multiplied by its utility, and the contributions from all outcome states are added. The utility scale makes explicit how the design treats the relative value of those states.

Choosing and Checking a Target

The feasible set is part of the definition of the target. It states which allocations the design team is willing or able to consider. For a finite set of candidate allocations, calculate \(W(\mathbf n)\) for each and select one with the largest value. For a continuous set of shares, the same principle is written as $$ \mathbf q^*=\underset{\mathbf q\in\mathcal Q}{\operatorname{arg\,max}}\ \sum_{j=1}^{J}q_j\mu_{U,j}, $$ where \(\mathcal Q\) is the specified feasible set of shares. In a real design, the set may reflect practical or scientific requirements. Those requirements should be stated rather than hidden inside the utility scores.

A useful diagnostic is to check the inputs before optimizing:

• For every arm, verify \(p_{jk}\geq 0\) and \(\sum_kp_{jk}=1\). A row that does not sum to one may indicate an omitted outcome state or a calculation error.
• Check that each utility score has the intended direction and that all arm scores refer to a comparable scale, perspective, and time horizon.
• Recalculate expected utility arm by arm. A very large contribution can be mathematically correct, but it is worth checking whether it comes from an unintended probability, score, or unit.
• State whether probabilities are assumptions or estimates. If they are estimated, the computed target inherits uncertainty from those estimates.

The principal alternative in the earlier material on Neyman allocation is an allocation chosen to improve the precision of a specified treatment-effect estimate, under its stated assumptions. That is a statistical information objective, whereas expected patient utility is a welfare objective. An allocation that is best for one objective need not be best for the other. A design may need to make the objective explicit before choosing a target; merely labeling an allocation “optimal” is incomplete.

This expected-utility formulation has limitations. It compresses several outcomes into a single weighted average, so different utility judgments can change the ranking of allocations. It also treats the input probabilities as adequate for the design question; uncertain or population-specific probabilities can make a numerical target fragile. Finally, maximizing this objective alone does not guarantee precise estimates, representativeness across arms, or the same benefit for every participant. Those properties require additional criteria, not an unspoken reinterpretation of \(W(\mathbf n)\).

Common Mistakes

• Using outcome probabilities without utilities. Probabilities describe what may happen; utilities specify how the outcomes enter the welfare objective. The objective requires both.
• Using utility scores without probabilities. A highly valued outcome contributes little to expected utility if its probability is small. Calculate the probability-weighted sum for each arm.
• Calling the result a precision optimum. Expected utility does not measure the variance of an estimator. Compare it with a Neyman target only after naming the different objective each addresses.
• Ignoring the feasible set. An allocation can be optimal among stated candidates without being optimal among all possible allocations. Report what was compared.
• Interpreting an expected total as a guaranteed outcome. \(W(\mathbf n)\) is an expectation under the assumed model, not a promise that the realized participants’ utilities will equal that number.