Allocation by Posterior Expected Utility

From Posterior Evidence to Allocation

An adaptive trial can use current posterior evidence to decide how likely the next participant is to receive each treatment arm. Posterior probability matching bases those chances on how often an arm is best across posterior draws. Allocation by posterior expected utility instead summarizes the clinical consequences of outcomes: an arm receives more allocation when its expected utility is higher, with the exact probabilities determined by a stated mapping from utility to allocation weight.

This distinction matters when outcomes have unequal clinical consequences. A modest increase in the chance of benefit may not compensate for a meaningful increase in serious harm. Expected utility gives the design a way to represent that trade-off explicitly, rather than treating every success or every chance of being best as equivalent.

Definition: Let $k$ index the currently eligible treatment arms, and let $Y_{\mathrm{new}}$ denote the potential clinical outcome for a participant assigned to arm $k$. Let $u_k(Y_{\mathrm{new}})$ be the prespecified utility of that outcome, on a common scale across arms. Given the data $D$, the posterior expected utility of arm $k$ is $$ \bar U_k=E\left[u_k(Y_{\mathrm{new}})\mid D,\text{ assignment to }k\right]. $$ An allocation rule maps these values to probabilities $\pi_k$ satisfying $\pi_k\geq 0$ and $\sum_k\pi_k=1$. The next participant is randomized using these probabilities.

The posterior expectation integrates uncertainty about the arm’s outcome distribution as well as the randomness of the future outcome. Its interpretation depends on the utility definition: a value measured in utility points is not automatically a probability or a clinical effect size. Utilities should be set before using the rule, with clinically meaningful benefits and harms represented on the same scale.

A Linear Utility-Weight Allocation Rule

Expected utilities can be negative, so directly dividing each $\bar U_k$ by their sum may produce negative or undefined probabilities. One explicit alternative is to convert utility to positive allocation weights using a common lower utility bound $L$ and a positive constant $c$:

w_k=c+\bar U_k-L,\qquad \pi_k=\frac{w_k}{\sum_j w_j}.

Here $L$ is a prespecified lower bound for the utility of an individual outcome under any eligible arm, and $c>0$ is a common baseline weight. Since $\bar U_k\geq L$, each weight is positive, so the resulting probabilities are valid. The baseline also gives every eligible arm positive allocation, provided the eligible set is nonempty. Both the lower bound and the baseline are part of the allocation policy, not quantities inferred automatically from the posterior.

This is one transparent mapping, not a uniquely mandated consequence of Bayesian decision theory. The utility scale and its mapping to randomization probabilities both matter. In particular, adding the same constant to every utility leaves the arm that maximizes expected utility unchanged, but it can change normalized allocation probabilities unless the reference $L$ is shifted consistently. A protocol should therefore specify the utility scale, reference, baseline, eligibility rules, and timing of updates.

Why the Probabilities Are Not a Utility Maximizer

The rule above creates a probability distribution from utility-based weights; it does not prove that randomizing according to those weights maximizes the expected utility of the next participant. To see the distinction, condition on the current data $D$. If the next participant is assigned to arm $k$ with probability $\pi_k$, the expected utility for that participant is

E[u_{\mathrm{new}}\mid D]=\sum_k\pi_k\bar U_k.

Because this is a weighted average of the arm-specific expected utilities, it is no larger than $\max_k\bar U_k$. If the sole objective were to maximize the next participant’s expected utility and there were no other constraints, assigning everyone to an arm with the largest $\bar U_k$ would achieve that maximum. A randomized rule instead trades off goals such as learning, sharing access, and avoiding complete concentration on an arm based on uncertain early evidence. Those goals need to be considered in design evaluation; they are not guaranteed by the normalization formula.

Worked Example: Binary Benefit and Harm With Unequal Weights

Consider two hypothetical arms, A and B. For each participant, benefit $R$ and harm $T$ are binary: each equals 1 if that outcome occurs and 0 otherwise. The prespecified utility is $u=10R-18T$, so benefit adds 10 utility points and harm subtracts 18. Assume the posterior model uses independent Beta(1,1) priors for each arm’s benefit and harm probabilities, with Bernoulli outcomes. Hypothetical data are: A has 8 benefits among 10 participants and 1 harm among 10; B has 6 benefits and no harms among 10.

With a Beta(1,1) prior, observing $x$ events in $n$ Bernoulli outcomes gives a Beta($1+x,1+n-x$) posterior, whose mean is $(1+x)/(2+n)$. Thus the posterior mean benefit and harm probabilities are $9/12=0.75$ and $2/12=1/6$ for A, and $7/12$ and $1/12$ for B. Since the utility is additive, its posterior expectation is 10 times the benefit probability minus 18 times the harm probability:

\bar U_A=10\left(\frac{9}{12}\right)-18\left(\frac{2}{12}\right) =7.5-3=4.5, \qquad \bar U_B=10\left(\frac{7}{12}\right)-18\left(\frac{1}{12}\right) =\frac{70}{12}-\frac{18}{12} =\frac{13}{3}\approx4.333.

The smallest possible individual utility is $-18$, when there is harm and no benefit, so take $L=-18$. Use $c=1$ utility point as the common baseline. Then $w_A=1+4.5-(-18)=23.5$, while $w_B=1+13/3-(-18)=70/3\approx23.333$. The allocation probabilities, using the exact fractions, are

\pi_A=\frac{47/2}{47/2+70/3} =\frac{141}{281}\approx0.5018, \qquad \pi_B=\frac{140/6}{281/6} =\frac{140}{281}\approx0.4982.

Thus the next participant has about a 50.18% chance of assignment to A and 49.82% to B under this particular rule. These are randomization probabilities conditional on the current posterior and the specified policy; they are not posterior probabilities that either arm is superior. The model assumptions used here are explicit: binary endpoints, the stated Beta priors, and the observed counts. Independence of benefit and harm is convenient for this example’s model, but the expected utility calculation itself follows by linearity of expectation and does not require the two outcomes to be independent.

Posterior Uncertainty and Probability Matching

Two arms can have the same posterior expected utility while having very different posterior uncertainty. Conversely, an arm can have a lower expected utility but a substantial probability of being the best. Expected-utility allocation retains the magnitude of utility differences; posterior probability matching retains the probability an arm is best, but not how large its advantage would be when best.

Worked Example: A Lower Chance of Being Best but Higher Expected Utility

Suppose hypothetical posterior summaries after observing trial data have the following form. Arm A’s utility is 8 with posterior probability 0.60 and $-4$ with probability 0.40. Arm B’s utility is 4 with probability 1. For the linear rule, set the common lower bound to $L=-4$ and the baseline to $c=1$.

The posterior expected utilities are

\bar U_A=0.60(8)+0.40(-4)=4.8-1.6=3.2, \qquad \bar U_B=1(4)=4.

Arm A is best with posterior probability 0.60, while B is best with probability 0.40. The utility-based weights are $w_A=1+3.2-(-4)=8.2$ and $w_B=1+4-(-4)=9$. Therefore

\pi_A=\frac{8.2}{8.2+9}=\frac{41}{86}\approx0.4767, \qquad \pi_B=\frac{9}{17.2}=\frac{45}{86}\approx0.5233.

The allocation gives B the larger probability because B has the larger posterior expected utility, even though A has the larger probability of being best. Posterior probability matching would allocate 0.60 to A and 0.40 to B in this example. Neither rule is a universal substitute for the other: they encode different ways to use posterior evidence.

Weights, Sensitivity, and Design Checks

Utility weights are consequential design choices, not technical decorations. Returning to the benefit–harm example, suppose the harm penalty is increased from 18 to 30 while the benefit weight remains 10. The utility is now $10R-30T$, its lower bound is $-30$, and retain $c=1$. Using the same posterior means, the weights become $1+7.5+30(1-1/6)=33.5$ for A and $1+10(7/12)+30(1-1/12)=34\frac{1}{3}$ for B. Consequently,

\pi_A=\frac{33.5}{33.5+34\frac{1}{3}} =\frac{201}{407}\approx0.4939, \qquad \pi_B=\frac{206}{407}\approx0.5061.

The change in harm weight moves the allocation from a slight preference for A to a slight preference for B. This is not a computational defect: the rule responds to the changed clinical trade-off. It is a reason to prespecify weights and examine whether plausible alternatives substantially alter allocation, trial-level safety, or the chance of assigning participants to an inferior arm.

Useful diagnostics include checking that posterior predictive outcomes are clinically plausible, verifying that all arms use the same utility scale and outcome horizon, and tracking how probabilities change after each update. If posterior estimates are dominated by prior assumptions, sparse data, or a poorly fitting outcome model, precise-looking allocation probabilities can be misleading. A probability vector should also be checked numerically: each entry must be nonnegative and the entries must sum to 1 up to rounding.

Allocation by posterior expected utility differs from two common alternatives. Posterior probability matching allocates in proportion to posterior probabilities of being best; it emphasizes rank uncertainty. Deterministic greedy allocation assigns the next participant to an arm with the largest posterior expected utility; it maximizes immediate expected utility when no other objective or constraint is included, but can concentrate allocation early. The linear rule here randomizes using utility-based weights. Other mappings are possible, but their properties and operating characteristics must be assessed rather than assumed.

Common Mistakes and Limitations

Normalizing raw expected utilities. If utilities are negative, $\bar U_k/\sum_j\bar U_j$ need not be a valid probability vector. Define a nonnegative weight transformation first.
Treating the mapping as uniquely determined. Posterior expected utility ranks arms, but does not uniquely prescribe randomization probabilities. State the reference, baseline, and transformation.
Interpreting allocation as evidence of superiority. A 60% allocation probability is a feature of the design rule, not a 60% posterior probability that the arm is best.
Ignoring utility sensitivity. If reasonable benefit or harm weights lead to very different allocations, report and evaluate that sensitivity before using the rule.

The linear rule is simple and interpretable, but its probabilities depend on the selected utility scale and reference. It does not by itself protect a control arm, cap allocation, handle delayed outcomes, or guarantee a desirable balance of learning and patient benefit. Such requirements must be built into and evaluated as part of the trial design.

Check Your Understanding

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

If two arms have posterior expected utilities 6 and 3, and the rule uses $L=-2$ and $c=1$, what are their allocation probabilities?
Why is dividing each arm’s expected utility by the sum of expected utilities not generally safe?
In a binary benefit–harm model with utility $8R-20T$, an arm has posterior mean benefit probability 0.70 and harm probability 0.10. What is its posterior expected utility?
Can an arm receive a higher allocation probability than another while having a lower posterior probability of being best? Explain.

Solutions. 1. The weights are $1+6-(-2)=9$ and $1+3-(-2)=6$, so the probabilities are $9/15=0.60$ and $6/15=0.40$. 2. Expected utilities may be negative, making the resulting ratios negative or the denominator zero; a specified nonnegative weight mapping is needed. 3. The expected utility is $8(0.70)-20(0.10)=5.6-2=3.6$ utility points. 4. Yes. Expected-utility allocation uses the magnitudes of utility across posterior possibilities, whereas probability matching uses the chance of being best; the two summaries can rank arms differently.