Why Pure Welfare Maximization Can Eliminate a Control Arm

When the Welfare Objective Has an Endpoint Solution

An allocation rule can be judged by what it achieves for the patients enrolled now, by how precisely it compares treatments, or by some combination of the two. These goals need not select the same allocation. Under pure welfare maximization, the question is narrowly defined: if each arm’s expected patient utility is known, which assignment gives the current cohort the greatest expected total utility? In a two-arm setting, if one arm has higher known expected utility for every eligible patient, the answer is to assign everyone to that arm.

This result is not a recommendation to omit a control group from a clinical trial. It is a consequence of optimizing an objective that counts current expected utility but assigns no value to comparative information. The distinction matters because a trial may need to learn whether the apparently better treatment really is better, and by how much. The earlier material on ethical allocation as expected patient utility motivates the welfare objective; here the focus is on what happens when that objective is optimized by itself.

Definition: Consider $N$ eligible patients and two arms, control $C$ and treatment $T$. Let $u_C$ and $u_T$ be their known expected utility per patient, on a common scale where larger values are better. Let $q$ be the fraction assigned to $T$, with $0 \le q \le 1$. Under an additive welfare objective, the expected total utility is $$ W(q)=N\bigl((1-q)u_C+q u_T\bigr). $$ Pure welfare maximization chooses $q$ to maximize $W(q)$, without including a value for learning about the treatment contrast.

The expression assumes that the planned cohort size is fixed, patients contribute additively to total utility, and the expected utility values apply to the patients being allocated. It also treats the two expected utilities as known inputs to the decision. The allocation is judged by its expected total, not by a guarantee about the realized outcomes of any particular cohort.

Why the Optimum Can Assign Everyone to One Arm

Expanding the objective makes its structure clear. The welfare change as the treatment share increases is constant: each additional unit of allocation share replaces control assignments with treatment assignments. Thus the objective is linear, not bowl-shaped or peaked in the interior.

W(q)=N u_C+Nq(u_T-u_C), \qquad \frac{dW(q)}{dq}=N(u_T-u_C).

Since $N$ is positive, the sign of the slope is the sign of $u_T-u_C$. If treatment has higher expected utility, increasing $q$ always increases expected total utility, so the maximum over $0\le q\le1$ occurs at $q=1$. If control has higher expected utility, the maximum occurs at $q=0$. If the expected utilities are equal, $W(q)$ is constant and every allocation share has the same expected total utility under this objective.

Result: Under the stated assumptions, pure welfare maximization gives $$ q^*= \begin{cases} 1, & u_T>u_C,\\ 0, & u_T<u_C,\\ \text{any value in }[0,1], & u_T=u_C. \end{cases} $$ When the expected utilities differ, this solution assigns no patients to the arm with lower expected utility. The result concerns the stated objective and assumptions; it does not establish that the assumed utility difference is true.

This endpoint solution follows directly from the slope. It is not a special feature of a particular numerical example. With fixed $N$, assigning a fraction $q$ to treatment changes expected total utility relative to assigning everyone to control by $Nq(u_T-u_C)$. If the difference is positive, any control assignments reduce expected utility relative to assigning those patients to treatment; if it is negative, the reverse holds.

Worked Example: A Known Utility Advantage for Treatment

Suppose a hypothetical trial plans to enroll $N=120$ patients. Assume the expected utility values are known and are $u_T=7.4$ utility units per patient for treatment and $u_C=5.9$ utility units per patient for control. Find the welfare-maximizing allocation and compare it with an equal allocation.

The expected utility difference is $u_T-u_C=7.4-5.9=1.5$ utility units per patient. Because this is positive, $W(q)$ increases with $q$, so the maximizing share is $q^*=1$. Assigning all 120 patients to treatment gives expected total utility $120(7.4)=888$ utility units.

With an equal allocation, 60 patients receive each arm. Its expected total utility is $60(5.9)+60(7.4)=354+444=798$ utility units. The difference is $888-798=90$ utility units, which also equals $60(1.5)=90$: the 60 control assignments each give up an expected 1.5 utility units relative to treatment under the assumed values.

These calculations identify the winner for the specified welfare objective. They do not show that the treatment truly has a 1.5-unit advantage, nor do they count any benefit from learning which arm is better.

The Comparative Information That Disappears

A treatment comparison requires observations from both arms, or other credible information about what would have happened under the unobserved arm. In a randomized trial, observations on both arms allow a direct comparison of their sample means. For example, if the outcome is measured in utility units, the usual difference in sample means is $\bar{x}_T-\bar{x}_C$, where each mean is calculated from patients assigned to that arm.

Under independent observations within and between arms, with finite arm-specific variances $\sigma_T^2$ and $\sigma_C^2$, the variance of this difference for fixed positive sample sizes $n_T$ and $n_C$ is $\sigma_T^2/n_T+\sigma_C^2/n_C$. This formula requires observations in both arms: each sample mean must exist. If $n_C=0$, there is no control sample mean to subtract, so the two-arm estimator is not defined. The problem is not merely that its standard error is large; the direct comparison cannot be calculated from the trial data.

Key distinction: The welfare objective evaluates expected outcomes for the current cohort using specified arm utilities. A comparative objective also values information about the difference between arms. If one arm has no participants, the trial cannot estimate the two-arm mean difference directly from its randomized observations.

The issue can also be seen as a lack of identification. Suppose everyone receives treatment, and the observed treatment outcomes are compatible with a treatment mean of 8 utility units. Those observations alone do not determine the control mean: a control mean of 7, 5, or another value could produce the same observed treatment data. Consequently, the treatment-control contrast could be positive, larger, or even negative without changing what was observed in the treatment-only trial. External evidence or additional assumptions could supply information about control, but the trial’s own data do not.

Worked Example: Precision Requires Observations in Both Arms

Consider a hypothetical planning calculation for a two-arm comparison with $N=100$. For illustration, suppose the outcome variances are $\sigma_T^2=9$ and $\sigma_C^2=4$, in squared utility units. Compare an allocation of 80 to treatment and 20 to control with an allocation of all 100 to treatment.

For 80 treatment and 20 control observations, the variance of the difference in sample means is $$ \operatorname{Var}(\bar{x}_T-\bar{x}_C) =\frac{9}{80}+\frac{4}{20} =0.1125+0.2000 =0.3125 $$ in squared utility units. Its standard error is $\sqrt{0.3125}=0.5590$ utility units.

For the all-treatment allocation, $n_T=100$ but $n_C=0$. The expression would require $4/0$, which is undefined; more fundamentally, $\bar{x}_C$ cannot be computed because no control outcomes were observed. Thus the trial cannot form this direct mean-difference estimate or its standard error.

This calculation illustrates an information loss, not a claim that the 80:20 allocation is best overall. Choosing an allocation by precision or by a combined objective requires a different criterion from pure welfare maximization.

Assumptions, Diagnostics, and the Limits of the Conclusion

The endpoint result is mathematically simple, but its interpretation depends on the assumptions behind the objective. In particular, the utility values must be treated as known and applicable to the patients in the planned cohort. If they are uncertain estimates, maximizing their estimated difference can select an arm that is not truly better. The algebra still gives an endpoint for the assumed values, but it does not account for uncertainty in those values.

The additive model also presumes that one patient’s assignment does not change another patient’s expected utility, and that the total objective is the sum of individual expected utilities. If treatment availability, capacity, or effects on other patients matter, this representation may not describe the relevant welfare. Likewise, if treatment effects differ across patients, a single pair of average utilities can conceal variation: an arm with the larger overall average need not have the larger expected utility for every patient.

A useful diagnostic is to ask what the objective counts and what it leaves out. Check whether both arms’ utility values are genuinely treated as known, whether the values apply to the patients being enrolled, and whether the objective includes the value of learning for future decisions. If the study is intended to estimate a treatment effect, also check that both arms receive assignments and that the design permits the comparison of interest. An allocation with zero probability of assignment to one arm provides no randomized observations from that arm.

Pure welfare maximization differs from allocation rules built to estimate a contrast efficiently. For example, Neyman allocation chooses sample-size shares to reduce the variance of a specified comparison, using the relevant outcome variances. Equal allocation is another common comparison design, though it is not necessarily optimal for every contrast or variance pattern. Neither rule follows from the pure welfare objective above: they answer a different question by giving weight to comparative information.

Allocation rationale	Primary quantity being optimized	Consequence for two-arm comparison
Pure welfare maximization	Expected utility for the enrolled cohort, using specified arm utilities	If one arm has greater expected utility, the optimum puts everyone on that arm; there is no direct comparison.
Precision-oriented allocation	Variance of a specified treatment contrast	Requires observations in both arms to estimate the contrast directly.
Equal allocation	A simple balanced design benchmark	Provides data in both arms, but balance alone does not guarantee minimum variance for every setting.

The central limitation is therefore not that the welfare calculation is incorrect. It is that a trial may have more than one purpose. If the aim includes learning whether treatment improves outcomes, an allocation that maximizes current expected utility while eliminating control observations leaves that aim unmet. The pure objective makes no trade-off between current welfare and information because information has no term in $W(q)$.

Worked Example: Equal Expected Utilities

Suppose a hypothetical cohort has $N=80$ patients and the specified expected utilities are $u_T=u_C=6.2$ utility units per patient. Find the welfare under treatment shares $q=0$, $q=0.5$, and $q=1$.

For any share $q$, $$ W(q)=80\bigl((1-q)6.2+q(6.2)\bigr) =80(6.2) =496 $$ utility units. In particular, $W(0)=496$, $W(0.5)=496$, and $W(1)=496$.

The welfare objective is indifferent among these allocations because its slope is $80(6.2-6.2)=0$. That indifference does not mean the allocations provide the same comparative information: an allocation with patients in both arms allows a direct comparison, whereas either endpoint does not.

Common Mistake

Wrong reasoning: “If treatment has the higher expected utility, assigning everyone to it is the best trial design, because it maximizes patient benefit.”

Correction: It maximizes the specified expected utility for the current cohort only if the utility values and model assumptions are accepted and information has no value in the objective. A trial designed to learn a treatment contrast has an additional purpose. With no control observations, the trial cannot estimate that contrast directly, so the welfare-maximizing allocation under known inputs need not be an adequate comparison design.

Key takeaway: With fixed cohort size and known expected utilities, pure welfare is linear in the treatment share. A strictly better arm therefore receives everyone in the mathematical optimum. The resulting absence of a control group is the cost of omitting comparative information from the objective, not evidence that information is unnecessary.

Check Your Understanding

Answer each question, then compare your reasoning with the solutions below.

For $N=50$, $u_C=4.8$, and $u_T=5.1$, which allocation maximizes pure welfare, and what is its expected total utility?
If $u_C=u_T$, does pure welfare maximization uniquely determine the allocation share? Explain.
Why can a trial assigning everyone to treatment not calculate the direct difference between the treatment and control sample means?
State one assumption or omitted objective that could make the pure-welfare conclusion inadequate as a trial-design recommendation.

Solutions. (1) Since $u_T-u_C=5.1-4.8=0.3>0$, the objective is increasing in $q$, so $q^*=1$. The expected total is $50(5.1)=255$ utility units. (2) No. The slope is $N(u_T-u_C)=0$, so every $q$ gives the same expected total under this objective. (3) With no control assignments, $n_C=0$, so there is no control sample mean $\bar{x}_C$; the direct difference $\bar{x}_T-\bar{x}_C$ is undefined. (4) For example, the recommendation may be inadequate if the study must learn the treatment-control contrast, because pure welfare assigns no value to that information. It may also be inadequate if the arm utilities are uncertain rather than known or do not apply to all enrolled patients.