Why Some Objectives Should Not Be Compensated
A dose may provide substantial benefit while also causing an unacceptable level of harm. A weighted utility score can make that dose appear attractive if its benefit is large enough to offset the harm numerically. But some clinical requirements are not preferences to trade away: a dose must provide at least a minimum level of efficacy, and its toxicity risk must remain below a safety limit. These requirements define which doses may be considered at all.
The distinction is between feasibility and preference. Feasibility asks whether a dose meets the minimum requirements. Preference determines which dose is best among those that do. Utility can still help compare acceptable doses, but it cannot make a dose acceptable if it violates a hard constraint.
The measures and thresholds must be defined in clinical terms. For example, \(E(d)\) might be a mean change from baseline in symptom score, with units of points, while \(Q(d)\) might be the probability of a dose-limiting toxicity, expressed as a proportion. A minimum efficacy of 5 points and a maximum toxicity probability of 0.15 are different kinds of quantities; each threshold must be compared with its matching measure. If efficacy is a response probability, the requirement might instead be \(E(d)\geq 0.40\).
Before applying the rule, specify the target population, endpoint definitions, follow-up period, candidate doses, and the source of \(E(d)\) and \(Q(d)\). The measures should refer to comparable populations and time horizons across doses. The requirements should also be stated before inspecting which dose wins, when practicable, to reduce the temptation to move a threshold to justify a preferred result. Monotonicity is not required by the definition: a dose can be assessed against both constraints even if efficacy or toxicity does not increase smoothly with dose.
Apply Feasibility Before Ranking
The sequence matters. For each dose, check both requirements; discard any dose that fails either one; then compare utility among the remaining doses. A dose with excellent efficacy but excessive toxicity is infeasible, just as a very safe dose with efficacy below the minimum is infeasible. Passing one criterion does not compensate for failing the other.
State \(E_{\min}\) and \(Q_{\max}\), including endpoint, units, population, and time horizon.
Mark a dose feasible only if \(E(d)\geq E_{\min}\) and \(Q(d)\leq Q_{\max}\).
Use utility or another stated preference criterion only within the feasible set.
State which doses were excluded, which requirement they failed, and how uncertainty was handled.
Worked Example: Selecting Among Doses That Meet Both Requirements
Suppose a hypothetical dose-finding analysis estimates mean symptom improvement \(E(d)\), in points, and the probability \(Q(d)\) of a specified toxicity during 12 weeks. The prespecified requirements are at least 50 points of improvement and a toxicity probability no greater than 0.18. Utility is a unitless preference score.
| Dose (mg) | \(E(d)\) (points) | \(Q(d)\) | \(U(d)\) |
|---|---|---|---|
| 10 | 38 | 0.06 | 52 |
| 20 | 50 | 0.10 | 66 |
| 30 | 58 | 0.17 | 74 |
| 40 | 62 | 0.25 | 82 |
The rule is $$ E(d)\geq 50\ \text{points},\qquad Q(d)\leq 0.18. $$ At 10 mg, \(38<50\), so the dose fails the efficacy requirement even though \(0.06\leq0.18\). At 20 mg, \(50\geq50\) and \(0.10\leq0.18\), so it is feasible. At 30 mg, \(58\geq50\) and \(0.17\leq0.18\), so it is also feasible. At 40 mg, efficacy meets its requirement, but \(0.25>0.18\), so the dose fails the safety constraint.
The feasible set is \(\mathcal{F}=\{20,30\}\) mg. Utility ranks these two doses: \(U(30)=74>66=U(20)\). Under the stated rule, 30 mg is selected. Its higher utility cannot bring 40 mg back into consideration, because 40 mg violates the hard toxicity limit.
Why a Penalty Is Not a Hard Constraint
A common alternative is to subtract a toxicity penalty from benefit or utility and maximize the resulting score. This can be useful when harms are tradeable preferences, but it does not generally enforce a maximum toxicity limit. Unless the penalty is constructed and bounded to guarantee exclusion beyond the limit, a sufficiently high benefit score may still win despite violating the requirement.
Worked Example: A Weighted Score Chooses an Infeasible Dose
Consider three hypothetical doses with a minimum efficacy of 50 points and a maximum toxicity probability of 0.12. Suppose the utility scores are already on a common preference scale. An analyst instead calculates a penalized score \(S(d)=U(d)-100Q(d)\), where \(Q(d)\) is a proportion.
| Dose (mg) | Efficacy (points) | \(Q(d)\) | \(U(d)\) |
|---|---|---|---|
| 10 | 44 | 0.04 | 46 |
| 20 | 55 | 0.10 | 62 |
| 30 | 61 | 0.16 | 80 |
The penalized scores are $$ S(10)=46-100(0.04)=42,\quad S(20)=62-100(0.10)=52,\quad S(30)=80-100(0.16)=64. $$ Maximizing \(S(d)\) selects 30 mg. However, 10 mg fails the efficacy requirement, and 30 mg has \(0.16>0.12\), so both are infeasible. Only 20 mg satisfies \(E(d)\geq50\) and \(Q(d)\leq0.12\). The constrained rule therefore selects 20 mg, regardless of the penalized score.
The calculation shows why the two formulations answer different questions. The penalty says how much toxicity is traded against utility on the chosen score scale. The hard constraint says that doses above a defined toxicity limit are not eligible. Increasing the penalty might change this example's ranking, but a penalty alone does not state or guarantee the required exclusion.
Handling Uncertainty at the Boundary
In practice, \(E(d)\) and \(Q(d)\) are usually estimated rather than known. A dose whose estimated toxicity is just below a limit may still have substantial uncertainty about whether its true toxicity exceeds that limit. Therefore, a decision rule should state whether feasibility is based on point estimates or on an uncertainty-aware criterion. The choice is consequential and should not be made implicitly after seeing the results.
One possible rule requires a lower confidence bound for efficacy to meet the minimum and an upper confidence bound for toxicity to meet the maximum. This is more conservative than checking point estimates alone. The confidence bounds and their coverage properties must be suitable for the decision. If both requirements are to hold jointly with a stated confidence level, the procedure should account for their joint uncertainty; two separate 95% bounds do not automatically provide 95% joint coverage.
Worked Example: Applying Simultaneous Bounds
Suppose the same requirements are \(E(d)\geq50\) points and \(Q(d)\leq0.15\). For this hypothetical analysis, the reported lower efficacy bounds and upper toxicity bounds are prespecified simultaneous one-sided 95% bounds for the two requirements. The dose estimates and bound margins are:
| Dose (mg) | \(\hat E(d)\) (points) | Efficacy margin (points) | \(\hat Q(d)\) | Toxicity margin | \(U(d)\) |
|---|---|---|---|---|---|
| 20 | 54 | 3 | 0.11 | 0.03 | 68 |
| 30 | 60 | 5 | 0.14 | 0.04 | 75 |
For this rule, calculate \(E_L(d)=\hat E(d)-\text{efficacy margin}\) and \(Q_U(d)=\hat Q(d)+\text{toxicity margin}\). At 20 mg, $$ E_L(20)=54-3=51\ \text{points},\qquad Q_U(20)=0.11+0.03=0.14. $$ Both \(51\geq50\) and \(0.14\leq0.15\), so 20 mg meets the uncertainty-aware requirements. At 30 mg, $$ E_L(30)=60-5=55\ \text{points},\qquad Q_U(30)=0.14+0.04=0.18. $$ Efficacy meets its requirement, but \(0.18>0.15\), so 30 mg does not meet the safety requirement under this rule. Although its utility score is higher, it is not eligible. The rule selects 20 mg among these candidates.
This conclusion is conditional on the stated bound procedure and its assumptions; it does not establish that the true toxicity probability is certainly below 0.15. The example illustrates how a conservative feasibility definition changes eligibility, not a universal prescription for which uncertainty rule to use.
Justification, Diagnostics, and Limitations
The constrained rule follows directly from the meaning of a noncompensatory requirement. Define feasibility by the conjunction of the two conditions. The decision problem is then to maximize utility over the doses satisfying that conjunction:
If a dose \(d'\) violates either condition, then \(d'\notin\mathcal{F}\), whatever its utility. Thus it cannot be the solution to this optimization problem. If \(\mathcal{F}\) contains one dose, that dose is the only eligible choice. If it contains several, utility can distinguish among them. If it is empty, the constrained optimization has no eligible solution; reporting the dose with the least violation as though it were feasible would change the rule.
Useful diagnostics include checking the units and definitions of both measures, examining estimates near each threshold, and confirming that the estimates concern the intended population and follow-up period. A reported feasible dose based only on rounded values may be misclassified near a boundary; make the decision using unrounded values where possible. Also inspect whether the model has adequate data support at each candidate dose. A precise-looking estimate does not remedy poor support or a misspecified model.
If no dose is feasible, report that result and identify the failed requirement for each candidate. Stakeholders may reconsider the clinical requirements, candidate doses, study design, or evidence base, but any revised rule should be explicit. Safety limits should not be relaxed merely because the preferred dose misses them. If feasibility depends on point estimates, explain that choice and its uncertainty; if it depends on bounds, specify the bound method and how joint uncertainty was addressed.
The main limitation is that a constraint is only as meaningful as its endpoint, threshold, and estimation procedure. A hard limit can create a sharp distinction between doses with nearly identical estimated risks on opposite sides of the boundary. That is not a mathematical error; it is a consequence of treating the threshold as nonnegotiable. It does make careful uncertainty assessment and transparent reporting especially important. This framework also does not determine the appropriate clinical threshold or utility scale: those require justified clinical and stakeholder input.
Common Mistake
Wrong reasoning: “The dose exceeds the toxicity limit, but its benefit is large enough to compensate, so its total score is highest.” This uses a compensatory preference calculation to override a requirement that was defined as hard.
Correction: Apply the efficacy and safety criteria first. A dose that fails either criterion is outside the feasible set, and utility is used only to rank the doses that remain. If the intended decision really allows benefit to trade against toxicity, then the objective is a preference tradeoff rather than a hard safety constraint, and it should be described as such.
Check Your Understanding
Answer each question using the stated decision rule. Full solutions follow the questions.
- A dose has efficacy 48 points and toxicity probability 0.03. The requirements are at least 50 efficacy points and toxicity no greater than 0.10. Is it feasible? Explain.
- Two doses meet both requirements. Their utility scores are 61 and 57. Which dose is selected under the constrained utility rule?
- A dose has utility 90 but toxicity probability 0.21 when the hard maximum is 0.20. Can the high utility make it eligible?
- A dose has estimated efficacy 53 points with a lower-bound margin of 4 points. The efficacy requirement is 50 points. Does it pass an efficacy rule based on the lower bound?
- What should be reported if every candidate dose fails at least one hard requirement?
Solutions. (1) No. Although \(0.03\leq0.10\), efficacy is \(48<50\), so the dose fails one required condition. (2) Select the dose with utility 61, because both are feasible and \(61>57\). (3) No. Since \(0.21>0.20\), the dose is infeasible regardless of its utility. (4) No. Its lower bound is \(53-4=49\) points, and \(49<50\). (5) Report that the feasible set is empty, state which requirement each candidate fails, and explain any uncertainty rule used. Do not describe an infeasible dose as satisfying the original decision rule.