The Welfare-Information Trade-Off Frontier

Why an Allocation Frontier Is Useful

A two-arm trial allocation target cannot generally be judged by a single criterion. Sending more patients to the arm with greater expected patient utility can raise the expected utility of participants, while moving away from the allocation that minimizes the variance of the treatment contrast can reduce precision. The welfare-information frontier describes the attainable combinations of these two objectives. It helps distinguish targets that represent a genuine trade-off from targets that are inferior on both objectives.

The frontier is a decision aid, not a rule that supplies a uniquely correct allocation. Its conclusions depend on the utilities, variances, trial size, and feasible allocation range used to construct it. All numerical values below are hypothetical.

Model, Assumptions, and Definitions

Consider a fixed-size, two-arm trial with \(n\) participants. Let \(r\) be the target share assigned to treatment arm 1, so the target share assigned to control arm 0 is \(1-r\). Assume \(r\) lies in a feasible interval \([a,b]\), where \(0<a\leq b<1\). The interval can encode requirements such as a minimum number of participants per arm. A continuous target share is used for the calculations; in an actual trial, allocation counts must be integers.

Let \(u_1\) and \(u_0\) be the expected patient utilities under assignment to arms 1 and 0, respectively, measured on the same utility scale. These expectations can incorporate treatment-related harms and costs when those have been included in the defined patient utility. Assume each participant’s expected utility depends on their assigned arm, and that the allocation does not change the expected utility for other participants. Under these assumptions, the expected utility per randomized participant is \(W(r)\). The expected total utility for the \(n\) participants is \(nW(r)\), so it gives the same ranking of targets.

For the contrast variance, suppose the outcome being compared has arm-specific variances \(\sigma_1^2>0\) and \(\sigma_0^2>0\), the trial uses fixed arm counts corresponding to the target shares, and outcomes are independent across participants. The estimator is the difference in sample means, \(\bar{x}_1-\bar{x}_0\). Under this model, its variance is

Thus \(W(r)\) is measured in utility units per participant, while \(V(r)\) is measured in squared outcome units. They should not be added or compared as if they shared a scale. The frontier instead displays the attainable pairs and makes the trade-off explicit.

Locating the Variance Minimum

The variance-minimizing share provides a reference point for the frontier. Differentiating the variance function with respect to \(r\) gives

Setting this derivative to zero requires \(\sigma_1/r=\sigma_0/(1-r)\), since all quantities are positive. Solving yields the unconstrained Neyman share. The second derivative is positive throughout \(0<r<1\), so this stationary point is the unique variance minimum. With a feasible interval, the minimizing share is the unconstrained value if it lies in the interval; otherwise it is the nearest endpoint.

Here \(r_N\) is the variance-minimizing share without allocation bounds, and \(r_V\) is the variance-minimizing share among feasible targets. This derivation concerns the variance curve; it does not establish that the variance-minimizing target also maximizes patient utility.

The Shape of the Welfare-Information Frontier

Write \(\Delta u=u_1-u_0\). Expected utility is linear in the allocation share:

The direction of this slope determines which side of the variance minimum can be efficient. If \(\Delta u>0\), raising \(r\) raises welfare. For any feasible \(r<r_V\), increasing the share toward \(r_V\) raises welfare and lowers variance, so that target is dominated. Every feasible target at or above \(r_V\) is efficient: moving to a larger share raises welfare but also raises variance, while moving to a smaller share lowers variance but also lowers welfare. If \(\Delta u<0\), the same reasoning reverses: targets above \(r_V\) are dominated, and the efficient targets lie at or below \(r_V\). If \(\Delta u=0\), welfare is constant, so the unique variance-minimizing feasible target dominates all other targets.

This result also explains how to read a plotted frontier. When \(u_1>u_0\), the efficient branch runs from the feasible variance minimum toward higher treatment shares: welfare rises along it, but so does contrast variance. Targets on the other branch are not trade-offs; they are worse in both dimensions than a target closer to the variance minimum. When \(u_1<u_0\), the efficient branch instead extends toward lower treatment shares.

Worked Example: Tracing an Unconstrained Frontier

Worked Example: Applying a Minimum Control Requirement

Worked Example: When Treatment Has Lower Expected Utility

Assumptions, Diagnostics, and Alternatives

The frontier is only as useful as the quantities used to construct it. Check that the utility values refer to the same population and utility scale, that the sign of \(u_1-u_0\) has been interpreted correctly, and that the variance formula matches the contrast and design being considered. In particular, the displayed formula is for a difference in independent arm means with fixed arm sizes. Dependence between outcomes, a different estimand, or a different design may require a different variance expression.

In practice, the arm variances may be estimated rather than known, and the utility values may also be uncertain. A frontier calculated from point estimates is then an estimated frontier, not a guarantee that the same ordering will hold under the true values. Useful checks include recalculating the efficient region across plausible inputs and checking whether the selected share respects the actual integer allocation and arm-size constraints. If small changes in an input move the variance minimum or reverse the utility ordering, report that sensitivity rather than presenting the frontier as exact.

A single-objective alternative is to choose the target that maximizes expected utility alone, or the one that minimizes contrast variance alone. These choices answer different questions: welfare-only maximization moves toward the arm with larger expected utility, while variance-only minimization selects \(r_V\). The frontier makes visible what each single-objective choice leaves aside. A decision-maker can then select a feasible point on the efficient branch using explicit study priorities or constraints. Without those priorities, the frontier does not determine a unique target.

Common Mistake