Why Response Probabilities Are Not the Whole Benefit
When patients can have different benefits from a response, the arm with the highest response probability need not provide the greatest expected benefit per patient. For example, a response might represent a small improvement under one treatment and a much larger improvement under another. An ethical allocation target should reflect the benefit being optimized, not response rates alone.
This tutorial considers a fixed number of patients allocated among arms. The objective is to maximize expected incremental patient benefit, subject to stated allocation requirements such as a minimum control share. It does not include treatment-related harm or cost. Benefit magnitudes must be on a common, meaningfully comparable scale, such as quality-adjusted life-year gains or a prespecified utility scale.
The product \(p_iB_i\) follows by averaging over the two response outcomes: a patient responds with probability \(p_i\) and then receives mean benefit \(B_i\); otherwise the incremental benefit is zero. If nonresponders can also receive incremental benefit, or if benefit is not adequately summarized by a response indicator and a responder mean, specify \(\mu_i\) directly as the arm’s expected incremental benefit instead. The allocation objective then uses those \(\mu_i\) values.
For the target calculations below, assume the benefit units are comparable across arms, the specified response probabilities and benefit magnitudes apply to the patients being randomized, and one patient’s assignment does not change another patient’s outcome. Assume a fixed total sample size and that the target’s welfare criterion is the sum of individual expected benefits. Treat the inputs as fixed for calculating the target; their uncertainty is a separate practical concern.
Allocation Constraints and the Welfare Objective
Write \(w_i\) for the target share in arm \(i\). Shares must sum to one. Let \(l_i\) be a required minimum share and \(u_i\) an optional maximum share, where \(0\leq l_i\leq u_i\leq1\). The feasible targets satisfy \(l_i\leq w_i\leq u_i\) and \(\sum_i w_i=1\). A minimum control group is one possible lower-bound constraint; it is not itself a benefit.
Because this objective is linear in the allocation shares, an unconstrained welfare optimum is generally at a boundary, not at an interior ratio proportional to response probabilities or benefits. If one arm has the greatest expected benefit, the target assigns it as much of the available allocation as the constraints permit. Arms with lower expected benefit receive their required minimums unless an upper bound on the best arm forces allocation elsewhere.
Why the Largest Expected-Benefit Arm Is Favored
The exchange argument gives a direct justification. Suppose a feasible allocation gives arm \(j\) some share above its minimum while arm \(i\) is not at its maximum, and \(\mu_i>\mu_j\). Shift a small amount \(\delta>0\) of allocation from \(j\) to \(i\). This keeps the total share unchanged and, if the bounds still hold, remains feasible. The change in expected benefit is positive:
Therefore, an allocation that leaves available capacity in a higher-benefit arm while placing discretionary share in a lower-benefit arm cannot be optimal. Repeating this exchange fills arms in order of decreasing \(\mu_i\), subject to their bounds. This argument also explains why the optimum can change abruptly when the ordering of expected benefits changes: a small change in an input can switch which arm receives the discretionary share.
Worked Example: Benefit Magnitudes Change the Target
Consider three arms in a hypothetical trial: control \(C\), treatment \(A\), and treatment \(B\). Suppose response probabilities are \(p_C=0.30\), \(p_A=0.50\), and \(p_B=0.70\). The mean incremental benefit among responders is \(B_C=1.0\), \(B_A=2.0\), and \(B_B=0.80\) utility units. Require at least 25% of patients in control and at least 10% in each treatment arm. There are no upper bounds below one.
Calculate expected benefit by arm. For control, \(\mu_C=p_CB_C=0.30(1.0)=0.30\) utility units per patient. For treatment \(A\), \(\mu_A=0.50(2.0)=1.00\). For treatment \(B\), \(\mu_B=0.70(0.80)=0.56\). Thus \(A\), not \(B\), has the greatest expected benefit, despite its lower response probability.
Find the target. The minimum shares use \(0.25+0.10+0.10=0.45\) of the allocation, leaving \(1-0.45=0.55\). Assign that remaining share to \(A\), the arm with the greatest \(\mu_i\). The target is \(w_C=0.25\), \(w_A=0.65\), and \(w_B=0.10\).
Check the objective. The target’s expected benefit is \(W=0.25(0.30)+0.65(1.00)+0.10(0.56)=0.075+0.650+0.056=0.781\) utility units per randomized patient. For comparison, assigning the remaining 0.55 to \(B\) would give shares \((0.25,0.10,0.65)\) and expected benefit \(0.25(0.30)+0.10(1.00)+0.65(0.56)=0.539\). Under the stated objective and constraints, the first target is better.
Benefit Differences Can Reverse Which Arm Leads
The ranking is determined by the products \(p_iB_i\), not by either factor in isolation. For two arms \(A\) and \(B\), arm \(A\) has the larger expected benefit exactly when \(p_AB_A>p_BB_B\). If response probabilities are positive, this is equivalent to \(B_A>(p_B/p_A)B_B\). This threshold makes explicit how much larger \(A\)’s responder benefit must be to outweigh a lower response probability.
Worked Example: A Benefit Threshold Reverses the Ranking
Suppose \(p_A=0.45\) and \(p_B=0.75\), with \(B_B=1.0\) utility unit per responder. The response-probability-only comparison favors \(B\), since \(0.75>0.45\). Find the threshold for \(B_A\).
The expected benefit in arm \(B\) is \(p_BB_B=0.75(1.0)=0.75\) utility units per patient. Arm \(A\) leads if \(0.45B_A>0.75\), or \(B_A>0.75/0.45=1.667\) utility units per responder. If \(B_A=2.0\), then \(\mu_A=0.45(2.0)=0.90\), so \(A\) leads. If \(B_A=1.5\), then \(\mu_A=0.45(1.5)=0.675\), so \(B\) leads.
To turn either comparison into a target, suppose control has a minimum share of 0.20 and each treatment has a minimum share of 0.10. When \(B_A=2.0\), the remaining share is \(1-0.20-0.10-0.10=0.60\), which goes to \(A\), giving target shares \((0.20,0.70,0.10)\) for \((C,A,B)\). When \(B_A=1.5\), it goes to \(B\), giving \((0.20,0.10,0.70)\). The change in target follows from the change in the expected-benefit ranking.
Upper Bounds and More Than One Competitive Arm
A cap can prevent the highest-benefit arm from receiving all discretionary allocation. In that case, the exchange argument still applies: fill the arm with the highest expected benefit up to its cap, then direct any remaining share to the best eligible arm. Continue until all shares are assigned. If two arms have the same expected benefit, reallocating share between them does not change \(W\), provided their constraints remain satisfied.
Worked Example: Allocating Under a Treatment Cap
In a hypothetical three-arm design, suppose the expected benefits are \(\mu_C=0.10\), \(\mu_A=0.80\), and \(\mu_B=0.60\) utility units per patient. The minimum shares are \(l_C=0.20\), \(l_A=0.10\), and \(l_B=0.10\). In addition, treatment \(A\) may receive at most \(u_A=0.60\); the other arms have no restrictive upper bounds.
The minimum shares total \(0.20+0.10+0.10=0.40\), leaving \(0.60\). Arm \(A\) has the highest expected benefit, but can receive only \(0.60-0.10=0.50\) more before reaching its cap. Assign 0.50 of the remainder to \(A\), then assign the last 0.10 to \(B\), the next-highest-benefit eligible arm. The target is \((w_C,w_A,w_B)=(0.20,0.60,0.20)\).
Its expected benefit is \(W=0.20(0.10)+0.60(0.80)+0.20(0.60)=0.02+0.48+0.12=0.62\) utility units per patient. The remaining share cannot be assigned to \(A\) without violating its cap. Assigning that 0.10 instead to control would reduce expected benefit by \(0.10(0.60-0.10)=0.05\), so \(B\) is the better feasible destination.
Assumptions, Diagnostics, and Alternatives
The target is only as meaningful as its inputs and objective. Check that all \(B_i\) values refer to the same benefit scale, response definitions are comparable across arms, and the stated responder means represent the population to be enrolled. If benefit varies substantially among responders, \(B_i\) should be the appropriate conditional mean for the objective; a single mean may conceal important patient-level differences. If nonresponders have nonzero benefit, replace \(p_iB_i\) with a correctly specified expected benefit for each arm.
Also verify that constraints match the design: minimum control share, treatment floors, and any caps should be explicit before optimizing. The calculated shares are continuous targets. With a finite trial, they may not correspond to integer sample sizes; rounding must respect the stated constraints and the total sample size. The target calculation itself does not establish that the specified probabilities or magnitudes are known precisely, nor does it account for uncertainty in them.
A useful diagnostic is to tabulate \(p_i\), \(B_i\), and \(\mu_i=p_iB_i\) separately. This reveals whether a surprising target results from a high response probability, a large benefit per response, or both. Check units and arithmetic, and perform a ranking check: if the target assigns discretionary share to an arm with lower \(\mu_i\) while a higher-\(\mu_i\) arm has capacity, either the target is not optimal for the stated objective or an unreported constraint has been applied.
The main alternatives optimize different quantities. Allocating by response probability alone is appropriate only if response probabilities themselves are the intended objective or responder benefits are equal across arms. Equal randomization can be useful for balance but does not generally maximize the stated expected-benefit objective. Neyman allocation is designed to improve precision for a specified statistical contrast; it need not maximize patient benefit. A precision requirement can constrain a welfare target, but the unconstrained welfare optimum described here should not be presented as if it also guarantees a desired level of information.
| Approach | Quantity used to rank or set allocation | Primary objective |
|---|---|---|
| Benefit-weighted ethical target | Expected benefit \(\mu_i=p_iB_i\) | Maximize expected patient benefit subject to stated bounds |
| Response-probability ranking | Response probability \(p_i\) | Favor arms with more responses, without weighting response magnitude |
| Neyman allocation | Outcome variability and contrast weights | Improve precision for a specified contrast |
| Equal allocation | Equal shares by design | Balance enrollment across arms, not necessarily maximize benefit |
This allocation rule is a welfare target under specified assumptions, not a universal ethical prescription. It does not include harm, cost, uncertainty in the benefit inputs, or a separate information requirement. Those considerations can alter the objective or feasible set and should not be smuggled into the interpretation of \(p_iB_i\).
Check Your Understanding
Answer each question, then compare your reasoning with the solutions below.
- Arm \(A\) has response probability 0.40 and mean benefit 3 utility units per responder. Arm \(B\) has response probability 0.60 and mean benefit 1.5 units. Which arm has greater expected benefit per patient?
- Three arms have expected benefits \(\mu_1=0.20\), \(\mu_2=0.70\), and \(\mu_3=0.50\). Their minimum shares are 0.20, 0.10, and 0.10, with no restrictive caps. What is the welfare-maximizing target?
- Why can’t one conclude that the highest-response-probability arm should receive the largest discretionary share?
- What should replace \(p_iB_i\) if nonresponders in arm \(i\) can also receive incremental benefit?
Solutions. (1) Arm \(A\) has \(0.40(3)=1.20\) utility units per patient; arm \(B\) has \(0.60(1.5)=0.90\). Arm \(A\) has greater expected benefit. (2) The minimum shares total 0.40, leaving 0.60. Assign that remainder to arm 2, the arm with the greatest expected benefit, giving target shares \((0.20,0.70,0.10)\). (3) Response probability omits the benefit magnitude among responders; the welfare ranking uses \(p_iB_i\), not \(p_i\) alone. (4) Use the arm’s expected incremental benefit per randomized patient, calculated to include benefits among both responders and nonresponders, and optimize the resulting weighted sum of allocation shares.