Why Rank Doses by Utility?
Two candidate doses can trade off efficacy against toxicity: one may achieve a higher probability of efficacy, while another may cause fewer adverse outcomes. A dose with the highest efficacy probability is not automatically preferred, just as a dose with the lowest toxicity probability is not automatically preferred. A decision requires a way to represent how important the possible benefits and harms are relative to one another.
Benefit-risk utility makes those consequences explicit. For each candidate dose, simulated patient outcomes are assigned utility values, and the dose is ranked by its average utility over the target population. The resulting recommendation is conditional on the outcome definitions, simulated population, time horizon, and values assigned to outcomes. Utility does not remove value judgments; it makes them visible enough to examine.
The probabilities can be estimated from simulated virtual patients generated under a specified pharmacokinetic/pharmacodynamic model, covariate distribution, and dosing regimen. In an empirical simulation with \(N\) patients, the expected utility is estimated by the mean of their assigned utilities:
This ranking is meaningful only if the simulation represents the population and treatment conditions relevant to the decision, the outcome definitions are applied consistently, and the utility values express the decision-maker's priorities. If the simulation compares doses using different populations, follow-up periods, or outcome definitions, their average utilities are not directly comparable.
Specify Outcomes and Values Before Ranking
For a simple example, let \(E=1\) indicate that a virtual patient meets a prespecified efficacy target over a specified period, and let \(T=1\) indicate a clinically important toxicity over that same decision period. Both are binary indicators: they equal 1 when the outcome occurs and 0 otherwise. One additive utility is
where \(w_E\) is the utility gain assigned to meeting the efficacy target, \(w_T\) is the utility loss assigned to toxicity, and \(c(d)\) is a dose-specific treatment burden or cost on the same utility scale. The weights and cost must be defined for the decision context; they are not estimated automatically by the PK/PD simulation. With this additive form,
This simplification follows from averaging the individual utility: \(E\) and \(T\) are indicators, so their means equal their probabilities, and \(c(d)\) is constant for patients receiving the same candidate dose. If the utility assigns different value to particular combinations of efficacy and toxicity, the relevant probabilities are joint probabilities instead. In either case, simulated efficacy and toxicity should be generated for the same virtual patients, so any dependence between the outcomes is retained.
Utilities may be expressed in abstract points, monetary units, quality-adjusted outcomes, or another explicitly defined scale. The scale itself is arbitrary up to a positive linear transformation when the only task is to rank expected utilities; the relative values assigned to outcomes are not arbitrary. A utility difference of 10 points has no clinical meaning unless the scale and its values have been explained.
Worked Example: Ranking Three Doses With Efficacy, Toxicity, and Burden
Suppose a hypothetical simulation produces 1,000 virtual patients per candidate dose. Efficacy success and clinically important toxicity are recorded jointly. The following counts are hypothetical; the four columns for each dose include every simulated patient.
| Dose | Efficacy and toxicity | Efficacy, no toxicity | No efficacy, toxicity | Neither | Dose burden \(c(d)\) |
|---|---|---|---|---|---|
| 100 mg | 30 | 570 | 20 | 380 | 0 points |
| 200 mg | 70 | 730 | 30 | 170 | 10 points |
| 300 mg | 120 | 800 | 80 | 0 | 20 points |
Assign \(w_E=100\) points to efficacy success and \(w_T=250\) points as the toxicity penalty. The utility for a patient is \(100E-250T-c(d)\). For 100 mg, there are \(30+570=600\) efficacy successes and \(30+20=50\) toxicities, so \(P(E=1\mid100\text{ mg})=0.600\) and \(P(T=1\mid100\text{ mg})=0.050\). Its estimated expected utility is
For 200 mg, the efficacy count is \(70+730=800\), and the toxicity count is \(70+30=100\). Thus,
For 300 mg, the efficacy count is \(120+800=920\), and the toxicity count is \(120+80=200\). Thus,
Under these specified values and this simulated population, the ranking is 100 mg (47.5 points), 200 mg (45.0 points), then 300 mg (22.0 points). The 100 mg dose has lower efficacy probability than the higher doses, but its lower toxicity probability and smaller burden produce the largest expected utility. This is a ranking under the stated utility, not a claim that 100 mg is best under every set of clinical priorities.
How Utility Weights Change the Recommendation
A recommendation should be examined for sensitivity to the values assigned to benefits and harms. In the previous example, hold \(w_E=100\) points and the burden values fixed, but let the toxicity penalty \(w_T=\lambda\) vary. The three expected utilities become
For example, if \(\lambda=100\) points, the utilities are 55, 60, and 52 points, respectively, so 200 mg ranks first. If \(\lambda=300\), they are 45, 40, and 12 points, respectively, so 100 mg ranks first. The change is not a contradiction: the decision criterion has changed because the cost assigned to toxicity has changed.
Worked Example: Finding a Weight Threshold
Using the same simulated probabilities, find the toxicity penalty at which 100 mg and 200 mg have equal expected utility. The difference between their utilities is determined by solving
Adding \(0.100\lambda\) to both sides gives \(60+0.050\lambda=70\). Subtracting 60 gives \(0.050\lambda=10\), and dividing by 0.050 gives \(\lambda=200\) points. At that value, each dose has expected utility 50 points. When \(\lambda<200\), 200 mg has the larger utility; when \(\lambda>200\), 100 mg has the larger utility. This threshold describes the comparison between these two candidates. It does not establish that either is preferred to every possible dose: 300 mg must also be compared at the chosen \(\lambda\).
Weight sensitivity can be reported as a range of recommendations rather than a single ranking. The range is useful when the utility values are uncertain, when stakeholders reasonably differ in their priorities, or when a recommendation changes under small changes in an assumed penalty. If a dose remains preferred over a clinically plausible range of weights, the decision is less dependent on one particular value judgment, though it is still conditional on the simulation and candidate set.
Why Joint Outcomes Can Matter
If utility is exactly additive in the efficacy and toxicity indicators, the two marginal probabilities are sufficient for its expected value. But many decisions use outcome values that are not additive. For example, toxicity may be especially undesirable when the efficacy target was not achieved, or a serious adverse event may have a different value from a mild event. In those cases, computing separate efficacy and toxicity probabilities and combining them without the joint outcome distribution can give the wrong result.
Worked Example: Equal Marginal Rates, Different Joint Outcomes
Consider two hypothetical doses, A and B, each simulated in 1,000 patients. Both have 700 efficacy successes and 100 toxicity events. Suppose the joint counts are as follows.
| Dose | Efficacy and toxicity | Efficacy, no toxicity | No efficacy, toxicity | Neither |
|---|---|---|---|---|
| A | 100 | 600 | 0 | 300 |
| B | 0 | 700 | 100 | 200 |
Assign 20 utility points to efficacy with toxicity, 100 points to efficacy without toxicity, -100 points to toxicity without efficacy, and 0 points to neither outcome. Dose A has estimated expected utility
Dose B has estimated expected utility
Both doses have efficacy probability 0.700 and toxicity probability 0.100, but A ranks above B under this joint-outcome utility. The reason is that A's toxicity events occur among patients who also meet the efficacy target, whereas B has 100 toxicity events without efficacy. The ranking follows from the specified outcome values; it is not implied by the marginal rates alone.
Assumptions, Diagnostics, and Limitations
Expected-utility ranking is a decision rule, not a guarantee that a dose will deliver the predicted outcomes in practice. Its use relies on several conditions:
- Relevant simulated population and regimen. The simulated covariate distribution, dosing schedule, and follow-up period should match the decision context. A shift in these conditions can change outcome probabilities.
- Well-defined outcomes. Efficacy and toxicity events need explicit definitions, thresholds, and time windows. They must be applied consistently across candidate doses.
- Credible joint simulation. When utilities depend on outcome combinations, the simulation must preserve clinically relevant dependence between efficacy and toxicity rather than combine unrelated marginal predictions.
- Defensible utility values. The assigned values should reflect the decision being made and be expressed on a common scale. A numerical ranking is not self-justifying.
- Comparable candidates and constraints. Candidate doses should be evaluated on a comparable basis. A hard safety or feasibility constraint should be applied as a constraint, rather than assumed to be captured by a utility penalty unless that is the intended decision rule.
Inspect the simulation before interpreting the ranking. For each dose, confirm that joint-outcome counts sum to \(N\), probabilities lie between 0 and 1, and any derived marginal probability agrees with the joint table. Check that the utility calculation uses the stated units and that a more harmful outcome receives a value no greater than a less harmful one when that ordering is intended. Investigate unexpected patterns in simulated exposure or outcomes against the underlying PK/PD assumptions; a utility calculation cannot correct a simulation that is implausible.
Monte Carlo error also matters. If \(s_U\) is the sample standard deviation of patient-level utility among the \(N\) simulated patients, the estimated standard error of the simulated mean is \(s_U/\sqrt{N}\), under independent sampling from the simulation distribution. A small difference in estimated expected utility may be unstable when it is comparable to simulation error. Increasing the number of simulated patients can reduce Monte Carlo error, but it does not resolve uncertainty in model parameters, outcome definitions, or utility values.
A common mistake is to rank doses by efficacy probability alone and call the dose with the largest value the best benefit-risk option. That reasoning ignores toxicity and any dose burden, so it answers a different question. The correction is to state the outcomes and utility rule first, calculate expected utility for each candidate using those same definitions, and examine whether reasonable alternative weights change the ranking. Another mistake is to treat utility points as an objective clinical fact. They encode chosen values; the probabilities may be model-based, but the values assigned to consequences require justification.
Compared with selecting a dose by a single efficacy target or a safety threshold, utility combines multiple consequences into one decision metric and can distinguish candidates that meet the same thresholds. Its trade-off is that the recommendation depends on utility values and can conceal clinically important details if only the final score is reported. Report the outcome probabilities alongside the ranking, identify the utility values used, and note any constraints. The ranking is also limited to the candidate doses that were evaluated; it does not prove that the highest-ranked candidate is optimal over all possible regimens.
Check Your Understanding
Answer each question using the definitions and decision rule above. Brief solutions follow the questions.
- A candidate dose has efficacy probability 0.75, toxicity probability 0.08, efficacy value 100 points, toxicity penalty 200 points, and burden 5 points. What is its expected utility under the additive rule?
- In a simulation of 500 patients, 40 have both efficacy and toxicity, and 260 have efficacy without toxicity. How many have efficacy success, and what is the estimated efficacy probability?
- Why can two doses with identical marginal efficacy and toxicity probabilities have different expected utilities?
- What does it mean if the preferred dose changes when the toxicity penalty is varied over a clinically plausible range?
- Does the dose with the greatest estimated expected utility necessarily satisfy a separate hard safety constraint? Explain.
Solutions. (1) \(100(0.75)-200(0.08)-5=75-16-5=54\) points. (2) There are \(40+260=300\) efficacy successes, so the estimated probability is \(300/500=0.600\). (3) If utility is non-additive, it can assign different values to the combinations of efficacy and toxicity; the joint outcome probabilities then affect expected utility even when the marginal probabilities match. (4) The recommendation depends on the chosen toxicity value, so the decision is sensitive to priorities; report the range of rankings and justify the values used. (5) No. A utility ranking does not automatically enforce a separate constraint. Apply the safety constraint explicitly and rank only among doses that satisfy it, if that is the decision rule.