Expected Utility of a Dosing Policy

Why Expected Utility Is a Policy-Level Quantity

A dosing policy is more than a dose for one patient: it is a rule that maps patient information to a dose. To compare two such rules, the evaluation must combine what each policy is expected to do clinically with how important those outcomes are to patients. Expected utility provides that combination. It expresses the average value of outcomes under a policy in a common scale, while allowing the value of benefit and harm to differ across patients.

The introductory material on quantifying clinical utility of dose recommendations established why clinical outcomes should be interpreted in terms of their consequences, not only their frequency. Here the focus is the calculation and aggregation: how to combine outcome probabilities with patient-level weights, and how to compare the resulting expected utility of competing policies.

Definition: Let $X$ denote measured patient characteristics, $A$ a dose or other dosing action, and $Y^a$ the clinical outcome that would occur for a patient if action $a$ were used. Let $u_x(y,a)$ be the utility assigned to outcome $y$ for a patient with characteristics $x$ under action $a$. For a policy $\pi$ that selects action $\pi(x)$, its expected utility in a target population is $$ U(\pi)=E\left[u_X\left(Y^{\pi(X)},\pi(X)\right)\right]. $$ The expectation averages over both the target population's patient characteristics and the outcomes that would occur under the policy.

Utility is a numerical representation of the relative value of possible outcomes. It may be defined directly in utility points or constructed from weights for clinical benefit and harm. Higher utility is preferred. The scale has no required physical unit, but its meaning must be stated: for example, utility points per treatment course. A utility value is not itself a probability, and a difference of one utility point has meaning only relative to the chosen scale and weights.

For a patient with characteristics $x$, the expected utility of a fixed action $a$ is $$ EU(a\mid x)=\sum_{y\in\mathcal{Y}}P(Y^a=y\mid X=x)u_x(y,a), $$ where $\mathcal{Y}$ is a set of mutually exclusive, exhaustive outcome categories. If outcomes are continuous, the corresponding expectation is an integral rather than a sum. The action-specific probability is about the outcome under action $a$, not merely the observed outcome among patients who happened to receive that action.

Deriving the Expected-Utility Calculation

For a discrete outcome, define $I(Y^a=y)$ to equal 1 when the potential outcome under action $a$ is $y$, and 0 otherwise. Exactly one indicator equals 1 for an exhaustive set of outcome categories. Therefore the realized utility can be written as a sum of outcome-specific utilities multiplied by indicators. Taking expectation conditional on the patient's characteristics gives the expected-utility formula.

\begin{aligned} E\left[u_x(Y^a,a)\mid X=x\right] &=E\left[\sum_{y\in\mathcal{Y}}I(Y^a=y)u_x(y,a)\mid X=x\right]\\ &=\sum_{y\in\mathcal{Y}}u_x(y,a)E\left[I(Y^a=y)\mid X=x\right]\\ &=\sum_{y\in\mathcal{Y}}P(Y^a=y\mid X=x)u_x(y,a). \end{aligned}

The conditioning remains on $X=x$ throughout because the random outcome in this derivation is the potential outcome $Y^a$. Equivalently, the probabilities can be written using an intervention: $P(Y=y\mid X=x,\operatorname{do}(A=a))$. Either notation describes the outcome distribution if action $a$ were applied, rather than confusing it with the observed association between action and outcome.

For a deterministic policy, the population expectation can be expanded by first calculating the conditional expected utility for each patient profile and then averaging over the target distribution of profiles:

U(\pi)=E_X\left[\sum_{y\in\mathcal{Y}}P(Y^{\pi(X)}=y\mid X)u_X(y,\pi(X))\right].

For a finite set of patient strata $s$, with target-population proportions $w_s$ that sum to 1, this becomes $U(\pi)=\sum_s w_s EU(\pi\mid s)$. The weights must represent the population in which the policy is intended to be used. A policy that performs well in a selected trial sample need not have the same overall utility in a population with different patient characteristics.

Worked Examples

Worked Example: Choosing Between Two Doses for One Patient

Suppose, hypothetically, that one patient has a target-treatment benefit worth $8$ utility points and a clinically important toxicity costs $12$ utility points. The three mutually exclusive outcomes are target attained without toxicity, target not attained without toxicity, and toxicity. Their utilities are respectively $+8$, $0$, and $-12$. Under a low dose, the model predicts probabilities $0.60$, $0.35$, and $0.05$. Under a high dose, it predicts $0.78$, $0.12$, and $0.10$. Each probability triplet sums to 1.

For the low dose, $$ EU(\text{low}\mid x)=(0.60)(8)+(0.35)(0)+(0.05)(-12)=4.80-0.60=4.20 $$ utility points per treatment course.

For the high dose, $$ EU(\text{high}\mid x)=(0.78)(8)+(0.12)(0)+(0.10)(-12)=6.24-1.20=5.04 $$ utility points per treatment course. With these stated probabilities and weights, the high dose has expected utility $5.04-4.20=0.84$ points greater for this patient. This is an expected-value comparison, not a guarantee that the high dose benefits this particular patient.

Worked Example: Aggregating Across Patient Strata

Consider a hypothetical target population with 40% of patients in stratum S and 60% in stratum R. Policy A and policy B select a dose for each stratum. In stratum S, target attainment is worth $10$ utility points and toxicity costs $20$; in stratum R, target attainment is worth $6$ points and toxicity costs $24$. In each stratum, the outcome categories are target without toxicity, toxicity, and neither. The hypothetical predicted probabilities and their implications are shown below.

Stratum and policy	Target without toxicity	Toxicity	Neither	Expected utility
S, policy A	0.70	0.05	0.25	6.00
S, policy B	0.80	0.08	0.12	6.40
R, policy A	0.50	0.10	0.40	0.60
R, policy B	0.75	0.05	0.20	3.30

For example, in stratum S under policy A, $EU=(0.70)(10)+(0.05)(-20)+(0.25)(0)=7.00-1.00=6.00$. In stratum R under policy B, $EU=(0.75)(6)+(0.05)(-24)+(0.20)(0)=4.50-1.20=3.30$. The other two entries follow in the same way.

Weighting the stratum results by their proportions gives $$ U(A)=(0.40)(6.00)+(0.60)(0.60)=2.40+0.36=2.76, $$ and $$ U(B)=(0.40)(6.40)+(0.60)(3.30)=2.56+1.98=4.54. $$ Thus policy B has $1.78$ more expected utility points per patient in this target population. It is important that both policies are averaged over the same stratum proportions; comparing unweighted stratum averages would answer a different question.

Worked Example: Averaging Across Predictive Scenarios

Suppose two plausible predictive scenarios are assigned weights $0.70$ and $0.30$. These weights might represent uncertainty about the model or parameter values; they are not the probabilities of the clinical outcomes themselves. For simplicity, use utility $+10$ for target attainment without toxicity, $-20$ for toxicity, and $0$ for neither. Under scenario 1, policy A predicts probabilities $0.75, 0.05, 0.20$, while policy B predicts $0.55, 0.05, 0.40$. Under scenario 2, policy A predicts $0.30, 0.20, 0.50$, while policy B predicts $0.65, 0.05, 0.30$. Each triplet sums to 1.

The scenario-specific utilities are: $$ \begin{aligned} EU(A\mid 1)&=(0.75)(10)+(0.05)(-20)=6.50,\\ EU(A\mid 2)&=(0.30)(10)+(0.20)(-20)=-1.00,\\ EU(B\mid 1)&=(0.55)(10)+(0.05)(-20)=4.50,\\ EU(B\mid 2)&=(0.65)(10)+(0.05)(-20)=5.50. \end{aligned} $$

Averaging over the scenario weights gives $$ U(A)=(0.70)(6.50)+(0.30)(-1.00)=4.25, \qquad U(B)=(0.70)(4.50)+(0.30)(5.50)=4.80. $$ On this weighted predictive average, policy B has $0.55$ more expected utility points. The result depends on the scenario weights and the utilities, so those inputs should be justified and examined rather than treated as unquestionable facts.

Assumptions and Diagnostics

The arithmetic is simple; the interpretation depends on whether its inputs are credible. First, the action must be sufficiently well-defined that “using dose $a$” has a consistent meaning. The outcome categories must be exhaustive and non-overlapping, and the probabilities for each patient profile and action must be between 0 and 1 and sum to 1. The utility assigned to each outcome should reflect the relevant patient-level valuation and the evaluation horizon.

To interpret the result causally, the probabilities must represent outcomes under the actions specified by the policy. In a randomized study, random assignment supports exchangeability of treatment groups, subject to appropriate handling of adherence and missing outcomes. In observational data, causal interpretation generally requires that, conditional on measured covariates, there be no unmeasured confounding of action and outcome. There must also be adequate positivity: for patient profiles where the policy recommends an action, there must be information supporting outcomes under that action. Consistency links the observed outcome to the potential outcome under the action actually received. These assumptions do not follow merely from calculating a model prediction.

Useful checks include verifying probability totals, tracing every utility weight to its clinical definition, checking that target-population weights sum to 1, and recalculating results under plausible alternative weights or model scenarios. Large changes in the preferred policy indicate that the choice is sensitive to assumptions. A single expected utility also hides the distribution of outcomes: a favorable average can coexist with a clinically unacceptable risk for some patients. It should not replace examination of important harms or a prespecified safety requirement.

Common Mistake and Alternatives

Common mistake: using observed probabilities $P(Y=y\mid X=x,A=a)$ as if they automatically described what would happen if every patient with profile $x$ received action $a$. This is not generally justified: patients receiving different doses may differ in prognosis, so the observed association can mix dose effects with patient selection. The correction is to use probabilities for the potential outcome under action $a$, $P(Y^a=y\mid X=x)$, or its intervention equivalent, and to state the assumptions or design that support those probabilities.

Expected utility is not the only useful summary. Reporting outcome probabilities alone makes clinical event rates visible but does not express the relative value of benefit and harm. A rule based on a dose threshold can be easier to communicate, but its choice depends on the harm tradeoffs that set the threshold; expected utility makes the weighted consequence explicit. Minimizing expected loss is mathematically equivalent when utility is defined as negative loss. A vector of separate benefit and harm summaries can also be informative when stakeholders do not agree on a single set of weights. None of these alternatives removes the need to specify the target population and the clinical consequences being considered.

The main limitation is that a numerical ranking can look more definitive than its inputs warrant. Patient preferences may vary, clinical effects may be uncertain, and an average can conceal heterogeneity. Expected utility is therefore best interpreted as a transparent summary conditional on stated probabilities, utility weights, and population—not as a universal ranking detached from those choices.

Check Your Understanding

Work through each question, then compare with the solutions below.

A dose has probabilities 0.50 for benefit, 0.10 for harm, and 0.40 for neither. Benefit is worth 6 utility points and harm costs 15. What is expected utility?
Why must action-specific outcome probabilities be distinguished from observed probabilities among patients who received that action?
Two patient strata have proportions 0.25 and 0.75. A policy has conditional expected utilities 8 and 2 points. What is its population expected utility?
What does it imply if the preferred policy changes when a plausible harm weight is used?

Solutions. (1) $EU=(0.50)(6)+(0.10)(-15)+(0.40)(0)=3.00-1.50=1.50$ utility points. (2) The treated and untreated groups may differ in prognosis; without suitable causal assumptions, observed outcome frequencies need not equal probabilities under an intervention assigning that action. (3) $U=(0.25)(8)+(0.75)(2)=2.00+1.50=3.50$ utility points per patient. (4) The ranking is sensitive to how harm is valued, so the weight should be justified and the sensitivity reported; the result does not establish a universally preferred policy.