Why Optimize Utility Across Doses?
A dose can improve clinical response while also increasing adverse effects or imposing greater treatment burden. A rule based only on response may therefore choose a dose that many patients would not prefer. Utility-based dose optimization makes the trade-off explicit: it assigns values to relevant outcomes, combines them into a dose-specific score, and selects the dose with the greatest expected score under the stated assumptions.
The score is not a property of the drug alone. It depends on which outcomes are included, how they are valued, and the population and time horizon to which the dose-response predictions apply. A defensible analysis therefore reports both the dose-specific predictions and the utility specification, rather than presenting the selected dose as an objective consequence of efficacy data alone.
Constructing a Dose-Specific Utility
Start by specifying the decision: the candidate doses, the population, the outcome horizon, and the outcomes that matter. Benefit might be a clinically meaningful response by a specified time; harm might be an adverse event over that same period; burden might reflect dosing frequency, monitoring, or another consequence of treatment. The definitions and measurement windows must be clear enough that the dose-specific probabilities refer to the same clinical question.
For a simple additive model, let \(R\) indicate a benefit outcome, \(A\) indicate a harm outcome, and \(b(d)\) represent treatment burden at dose \(d\). One possible utility is
Here \(v_R\) is the utility value assigned to benefit, \(v_A\) is the penalty assigned to harm, and all terms must be expressed on a common utility scale. The expected-utility formula follows from averaging the additive utility over possible outcomes. For example, \(E[v_RR\mid d]=v_RP(R=1\mid d)\) because \(R\) is an indicator, so \(E[R\mid d]=P(R=1\mid d)\); the same reasoning applies to \(A\). The burden term is fixed at dose \(d\) in this formulation.
This construction assumes that the selected outcome definitions and preferences are appropriate for the decision, that the dose-specific probabilities are relevant to the target population and horizon, and that the additive utility structure is acceptable. Additivity means that the value of benefit does not change according to whether harm also occurred, and vice versa. This is a modeling choice, not a universal fact. If preferences depend on combinations of outcomes, the utility must assign values to the combined outcome profiles instead.
Utility points are not automatically clinical units. They may come from elicited patient preferences, a prespecified decision framework, or a transparent scoring convention. Whichever source is used, the values and their rationale should be reported. If a score is only a pragmatic ranking scale, it should not be described as a validated measure of patient preference.
Worked Example: Comparing Three Candidate Doses
Suppose a hypothetical dose-ranging analysis predicts the following probabilities of meaningful response and adverse event over a common 12-week horizon. For illustration, assign 100 utility points to response, a 40-point penalty to an adverse event, and a burden cost of 0.2 points per milligram of daily dose. These values are hypothetical, and the dose burden score is part of the illustrative scoring rule, not a clinical measurement.
| Dose (mg/day) | Response probability | Adverse-event probability | Burden cost (points) |
|---|---|---|---|
| 5 | 0.45 | 0.05 | 1 |
| 10 | 0.65 | 0.08 | 2 |
| 20 | 0.78 | 0.16 | 4 |
| 40 | 0.82 | 0.30 | 8 |
The dose-specific expected utility is \(EU(d)=100P(R=1\mid d)-40P(A=1\mid d)-0.2d\). At 5 mg/day,
At 10 mg/day, \(EU(10)=100(0.65)-40(0.08)-2=65-3.2-2=59.8\) points. At 20 mg/day, \(EU(20)=78-6.4-4=67.6\) points. At 40 mg/day, \(EU(40)=82-12-8=62\) points. Thus 20 mg/day has the highest expected utility among these four candidates under the stated scoring rule. This conclusion is conditional on the specified probabilities, horizon, and weights; it does not establish that 20 mg/day is best under every reasonable set of preferences.
Weights Can Change the Decision
A useful check is to ask how much the preferred dose depends on the assigned values. Consider the 10 mg/day and 20 mg/day candidates in the example above, but let \(w\) be the adverse-event penalty in utility points, holding the response value at 100 and the burden costs fixed. The expected-utility difference is
The doses have equal expected utility when \(11-0.08w=0\), or \(w=137.5\) points. If the adverse-event penalty is below 137.5 points, 20 mg/day has the higher expected utility; if it is above 137.5 points, 10 mg/day has the higher expected utility. This is a break-even calculation, not a claim that one penalty is correct. It shows which preference assumption is driving this comparison and gives decision-makers a concrete quantity to discuss.
The same principle applies to benefit weights and treatment burden. A reported result is more informative when it identifies whether the selected dose remains preferred across plausible values, rather than giving only the dose for one arbitrary score. Sensitivity analysis that systematically varies weights belongs to a broader assessment, but even a simple break-even calculation can reveal a fragile choice.
Worked Example: A Different Harm Preference
Use the same hypothetical predictions for 10 mg/day and 20 mg/day, but suppose the decision framework assigns a 150-point penalty to an adverse event. The response value remains 100 points, and burden costs remain 2 and 4 points, respectively.
For 10 mg/day, \(EU(10)=100(0.65)-150(0.08)-2=65-12-2=51\) points. For 20 mg/day, \(EU(20)=100(0.78)-150(0.16)-4=78-24-4=50\) points. The preferred dose under this weighting is 10 mg/day, by 1 utility point. This is consistent with the break-even result because \(150>137.5\). The small difference also signals that the ranking is close under this scoring rule; it should not be communicated as a large or certain clinical advantage.
Outcome Profiles and Nonlinear Preferences
The binary, additive formula is convenient but may be too simple. Patients may distinguish mild from severe benefit, or mild from serious adverse events. A dose may also have treatment burden that is not well represented by a linear cost per milligram. In those cases, define a set of mutually exclusive outcome profiles \(k\), assign a utility \(u_k\) to each profile, and compute
This expression accommodates nonlinear preferences because each profile receives its own value. The profiles must be sufficiently complete for the decision: if a material outcome is omitted, its value cannot affect the result. Probabilities must also describe the same population, follow-up horizon, and definitions used to assign the utility values.
Worked Example: Valuing Different Response Levels
Consider two hypothetical doses and three mutually exclusive response categories at 12 weeks: no meaningful response, moderate response, and major response. Assign utilities of 0, 50, and 80 points to these categories. For simplicity, use a 40-point adverse-event penalty and the same burden costs as before: 2 points at the lower dose and 4 points at the higher dose.
| Dose | No response | Moderate response | Major response | Adverse-event probability |
|---|---|---|---|---|
| 10 mg/day | 0.20 | 0.50 | 0.30 | 0.08 |
| 20 mg/day | 0.10 | 0.40 | 0.50 | 0.16 |
For 10 mg/day, the response utility is \(0(0.20)+50(0.50)+80(0.30)=0+25+24=49\) points. Subtracting the expected adverse-event penalty and burden gives \(49-40(0.08)-2=49-3.2-2=43.8\) points. For 20 mg/day, response utility is \(0(0.10)+50(0.40)+80(0.50)=0+20+40=60\) points, so total expected utility is \(60-40(0.16)-4=60-6.4-4=49.6\) points. Under these hypothetical values, 20 mg/day is preferred by 5.8 points. This calculation treats response-category utility and adverse-event penalty additively; if the value of a response differs for patients experiencing an adverse event, combined profiles need separate utilities.
Assumptions, Diagnostics, and Alternatives
Utility maximization is only as credible as its inputs and structure. Before interpreting a result, check that the dose-specific probabilities are estimated for the intended population and outcome horizon; that outcome categories are clearly defined; that utilities use a coherent scale; and that no important benefit, harm, or burden has been omitted. If utilities are elicited from different sources, confirm that their scales can legitimately be combined. If probabilities do not sum to one in a profile-based analysis, the profiles may be incomplete or the inputs inconsistent.
A practical diagnostic is to calculate each component separately and then verify the total. Inspect whether the dose ranking is driven mostly by a single weight, a single outcome probability, or an unusual burden score. Check that increasing a penalty for a harm does not inadvertently increase expected utility, and that adding a treatment burden cost cannot improve a dose's score. These directional checks can expose sign errors or unit mismatches. Also compare the result with the component predictions: a surprising selected dose may reflect an unintended weight or coding error rather than a meaningful clinical trade-off.
Utility-based selection differs from selecting a dose by target attainment, which asks whether a specified response target is achieved with sufficient probability. A target rule can be clear and clinically interpretable, but it may not distinguish doses that both meet the target or represent harms and burden in the same score. An efficacy-only rule ranks doses by benefit and can be useful when the decision is specifically about efficacy; it does not by itself resolve trade-offs with harms or burden. Utility analysis combines valued consequences into one decision criterion, but requires explicit value judgments that those alternatives may not require.
The selected dose is the maximizer only among the candidate doses evaluated. It is not automatically an optimum over all possible doses, and a score difference is not a confidence statement or proof of clinical superiority. The calculation also does not, by itself, account for uncertainty in estimated dose-response probabilities, variation in preferences across patients, or model uncertainty. Those are important limitations; uncertainty and heterogeneity should not be hidden by reporting a single expected-utility ranking.
Applying the Decision Rule
Define candidate doses, target population, outcome horizon, and clinically relevant benefit, harm, and burden outcomes.
Assign values to outcomes using a stated preference source or transparent scoring convention, and make sure all terms share a common scale.
Combine dose-specific outcome probabilities with the corresponding utility values, checking arithmetic and probability totals.
Select a maximizer among the evaluated doses, then report the assumptions, component contributions, and any important sensitivity of the ranking to weights.
The result is a decision aid, not a substitute for clinical judgment. Its main value is that it makes trade-offs inspectable: readers can see what the model predicts, how outcomes were valued, and which assumptions lead to the selected dose.
Check Your Understanding
Answer each question using the definitions and decision rules above. Brief solutions are included below the questions.
- For \(EU(d)=100P(R=1\mid d)-40P(A=1\mid d)-b(d)\), calculate expected utility when the response probability is 0.70, adverse-event probability is 0.10, and burden cost is 3 points.
- In the 10 versus 20 mg/day comparison with response value 100, burden costs 2 and 4, and adverse-event probability difference 0.08, what harm penalty makes the doses tie?
- Why must outcome probabilities refer to a common time horizon when calculating expected utility?
- What assumption is made when response utility and adverse-event penalty are added separately, and what can be done if that assumption is unsuitable?
- Does the dose with the highest expected utility necessarily have the greatest efficacy? Explain.
Solutions. (1) \(100(0.70)-40(0.10)-3=70-4-3=63\) utility points. (2) The difference is \(11-0.08w\), so the tie occurs at \(w=11/0.08=137.5\) points. (3) Probabilities measured over different horizons refer to different chances of outcomes, so combining them would not describe one coherent decision. (4) Separate addition assumes the value of response does not change depending on whether an adverse event occurs, and vice versa. If that is unsuitable, assign utilities to combined outcome profiles. (5) No. A dose with less efficacy can have greater expected utility if its harm and burden costs are sufficiently lower under the specified utility function.