Decisions Require More Than Posterior Probabilities
A posterior probability can summarize evidence about an arm, but a probability by itself does not say which action is best. Selecting an arm may bring benefit, expose patients to harm, and require resources; stopping may avoid further costs but forgo a useful treatment; continuing may preserve the option to learn more while incurring its own cost. To compare these actions, each needs a utility on a common scale.
The decision is made using the posterior distribution after the data observed so far. For each feasible action, average its utility over the unknown treatment effects according to that posterior distribution. The action with the greatest posterior expected utility is preferred, provided the utilities adequately represent the decision-maker's objectives.
In a multi-arm trial, an action might be to stop, select one of \(K\) treatment arms, or continue under a specified plan. The utility function must state how outcomes, harms, and costs count for each action. Utilities can be in any consistent units, but the same scale and perspective must be used for every action being compared. A cost expressed in utility units is subtracted; it is not a probability threshold.
From Posterior Uncertainty to a Decision
The calculation has two layers. First, specify the utility that would result for each action and state. Second, average that utility over the posterior distribution. The averaging is essential: choosing the arm with the largest estimated effect, or the largest probability of benefit, is not generally the same as choosing the arm with the greatest expected utility.
Here \(\mathcal{A}\) is the set of feasible actions and \(\mathcal{S}\) is a set of mutually exclusive, exhaustive states. The sum form makes the logic explicit: each state-specific utility is weighted by its posterior probability. For a continuous treatment effect, the integral form in the definition does the same job.
Include only actions that can actually be taken at this decision point, such as stop, select arm \(j\), or continue under a defined plan.
Specify the consequences and costs for each action on one scale. If benefit and harm can occur together, make clear how their utilities combine.
Multiply state-specific utilities by their posterior probabilities, or integrate utility over the posterior distribution.
Choose an action with the largest posterior expected utility. If actions tie, the utility calculation alone does not distinguish them.
For a continuation action, its utility must represent what continuing means. If a continuation plan includes future observations and a later action, its state-specific utility can be defined as the expected value of that whole plan conditional on the current state, including its costs. The current decision then averages that value over the current posterior. This tutorial uses stipulated continuation values in an example; deriving their future-data distribution is a separate calculation.
When a Probability Cutoff Is a Utility Rule
A fixed probability cutoff can be justified as an expected-utility rule in a simple special case. Suppose selecting arm \(j\) yields \(R\) utility units if it achieves a prespecified benefit, yields zero otherwise, and incurs a fixed cost \(C\). Stopping has utility zero. If \(p_j=P(\text{benefit for arm }j\mid D)\), then
Selecting the arm is preferred to stopping precisely when \(R p_j-C>0\), or \(p_j>C/R\), assuming \(R>0\). Thus the cutoff follows from the reward and cost in this particular utility model. It is not a universal constant. If harms have a separate penalty, costs vary by arm, or utilities differ by outcome severity, a cutoff on benefit probability alone may not represent the decision.
Worked Example: A Cutoff and Expected Utility Give Different Decisions
Suppose two hypothetical arms have posterior treatment effects \( \theta_A\mid D\sim N(0.56,0.10^2)\) and \( \theta_C\mid D\sim N(0.55,0.10^2)\), measured as absolute improvement in a clinical score. A clinically meaningful benefit is an effect greater than \(0.50\) score points. Selecting either arm gives 5 utility points if that benefit is achieved and 0 otherwise, and incurs a fixed selection cost of 0.50 utility points. Stopping has utility 0. Assume this is the complete utility model for this example.
For arm A, the posterior probability of benefit is \(P(\theta_A>0.50\mid D)=\Phi((0.56-0.50)/0.10)=\Phi(0.60)\approx0.7257469\). Its expected utility is calculated using the unrounded probability:
For arm C, \(P(\theta_C>0.50\mid D)=\Phi((0.55-0.50)/0.10)=\Phi(0.50)\approx0.6914625\), so
Stopping has expected utility 0, so arm A is preferred. Now compare this with a fixed rule that selects an arm only if its posterior benefit probability exceeds 0.75. Neither arm passes that cutoff, so the rule would stop. The discrepancy is not a mathematical contradiction: under the stated utility model, the benefit probability needed to beat stopping is only \(0.50/5=0.10\). The 0.75 cutoff encodes a different decision preference.
Accounting for Benefit, Harm, and Arm-Specific Cost
When benefit and harm both matter, the posterior probability of benefit alone is insufficient. The joint posterior distribution is useful for describing which outcomes occur together, but expected utility can be computed from marginal probabilities when the utility is additive. No independence assumption is needed for that addition: expectation is linear.
Worked Example: Selecting Between Arms With Benefit and Harm Penalties
Consider hypothetical posterior classifications for two arms. Let benefit mean clinically meaningful improvement and harm mean clinically important toxicity. Suppose selecting an arm gives \(+6\) utility units for benefit, \(-8\) for harm, and incurs a fixed cost of 1 utility unit. If benefit and harm both occur, both outcome utilities apply. Stopping has utility 0. The posterior probabilities are:
| Arm | \(P(\text{benefit}\mid D)\) | \(P(\text{harm}\mid D)\) |
|---|---|---|
| A | 0.60 | 0.15 |
| B | 0.50 | 0.05 |
For either arm \(j\), the utility is \(6I_B-8I_H-1\), where \(I_B\) and \(I_H\) indicate benefit and harm. Taking the posterior expectation gives \(6P(B\mid D)-8P(H\mid D)-1\). For arm A:
For arm B:
Arm B is preferred to arm A and to stopping, whose utility is 0. Arm A has the higher probability of benefit, but its greater posterior probability of harm lowers its expected utility enough to change the ranking. The calculation did not assume benefit and harm were independent; additive utilities allow their marginal expectations to be combined.
Including the Value of Continuing
Continuing is a distinct action, not a synonym for selecting an arm now. Its value depends on the specified continuation plan, including any future decision and the costs or burdens incurred along the way. In some settings, continuing has greater expected utility because it preserves options or generates information; in others, its costs make stopping or selecting immediately preferable.
Worked Example: Comparing Stop, Select, and Continue
Suppose a hypothetical trial has three mutually exclusive posterior states for its leading candidate: favorable, inconclusive, and harmful, with probabilities 0.60, 0.30, and 0.10. These sum to 1, so the listed states are exhaustive for this simplified decision. The utilities below are stipulated on a common scale. Selecting now yields 8 units in the favorable state, \(-1\) in the inconclusive state, and \(-12\) in the harmful state. Stopping yields 0 in every state. Continuing under a specified plan has state-specific utilities 5, 4, and 0, respectively; these values already include the plan's costs and downstream consequences.
The expected utility of selecting now is
For stopping, the expected utility is \(0.60(0)+0.30(0)+0.10(0)=0\). For continuing, it is
Continuing is preferred because \(4.20>3.30>0\). This conclusion depends on the stipulated continuation utilities, not merely on the favorable-state probability. In an actual analysis, those utilities need a defensible basis and must reflect the proposed continuation plan rather than an imagined ability to know the true state.
Derivation and Checks for a Continuation Value
The averaging rule follows from conditioning on the unknown state. Let \(Y\) denote future data under a continuation plan and let \(V(Y,\theta)\) be the utility delivered by the plan after those data are observed and its prescribed later action is taken. If the plan has a cost \(c\), its utility conditional on the current state can be written
Averaging this conditional value over the current posterior gives
The equality is the law of iterated expectation: first average the plan's later value over future data conditional on the state, then average over current uncertainty about the state. This decomposition also identifies a common error. Assigning continuing the payoff it would receive if the true state were known, without accounting for the future data and prescribed decision, overstates what the plan can deliver.
Diagnostics, Alternatives, and Limitations
Before trusting a maximizing action, check the decision calculation as carefully as the posterior calculation. A useful audit includes:
- Check the action set. Make sure the comparison includes the feasible alternatives, including stopping when it is available.
- Check the utility scale. Confirm that benefit, harm, and cost are expressed from the same perspective and in compatible units.
- Check posterior coverage. State categories in a discrete calculation must be exhaustive and mutually exclusive; probabilities should sum to 1, apart from rounding.
- Check signs and costs. A harm penalty or cost should reduce utility under the stated convention. Recalculate if a sign change alters the selected action.
- Check sensitivity. If small plausible changes in a utility or posterior probability change the maximizing action, report that dependence rather than presenting the choice as inevitable.
The main alternatives answer different questions. A fixed posterior-probability threshold is simple and may be appropriate when its implied rewards and costs are acceptable. Choosing the arm with the highest posterior mean or the greatest probability of benefit ranks a particular feature of the posterior, but does not account for every consequence. Maximizing posterior expected utility combines the specified consequences, yet it is only as credible as the utility model and posterior distribution supplied to it.
Expected utility has limitations. It does not remove value judgments: the utility assigned to benefit, harm, cost, or delay must be chosen and defended. A single average can also conceal substantial posterior uncertainty or sensitivity to rare but serious harms. The calculation does not guarantee that the posterior model is adequate, that all relevant outcomes were included, or that the available actions are feasible. It makes assumptions in the decision explicit; it cannot make poor assumptions harmless.
Check Your Understanding
Work each question, then compare with the brief solutions below.
- An arm gives \(+4\) utility units if benefit occurs, 0 otherwise, and has a fixed cost of 1 unit. Stopping yields 0. What posterior benefit probability makes selecting the arm break even?
- Two actions have posterior expected utilities 2.1 and 2.1. What can be concluded from expected utility alone?
- Why is it not generally valid to select the arm with the largest posterior benefit probability when harms and costs differ among arms?
- A continuation plan's conditional utility is calculated assuming the true state will be known before its later action. What is the problem, and what should the calculation represent instead?
Solutions. (1) Set \(4p-1=0\), giving \(p=0.25\); selection exceeds stopping when \(p>0.25\). (2) The actions tie under the specified utility model; expected utility alone does not select between them. (3) Benefit probability omits other utility-relevant consequences. Different harm probabilities, harm penalties, or costs can reverse the ranking, so compare the full posterior expected utilities. (4) The assumed perfect knowledge can overstate the plan's value. Its conditional utility should average the later outcome and action over future data under the plan, then be averaged over the current posterior.