Why Dominance Matters in Dose Selection
A dose decision can involve several benefits and several harms. One dose might improve a clinical outcome more while also causing more toxicity; deciding between those doses requires a preference rule. But sometimes a comparison is simpler: one dose provides at least as much of every valued benefit and no more of any valued harm. In that case, no trade-off is needed to rule out the inferior option under ordinary benefit-risk preferences.
The result is useful precisely because it does not require choosing utility weights or specifying how much benefit compensates for a particular harm. Its guarantee is conditional, however: outcomes must be compared on the same dimensions, and the utility must respect the stated directions of preference. Also, dominance rules out a lower utility for the dominating dose; it does not always guarantee a strict preference.
The Utility Guarantee
Write \(U(\mathbf B,\mathbf H)\) for a utility function of the benefit vector \(\mathbf B\) and harm vector \(\mathbf H\). Assume \(U\) is nondecreasing in each benefit while the other outcomes are held fixed, and nonincreasing in each harm while the other outcomes are held fixed. In plain terms, increasing one benefit cannot lower utility, and increasing one harm cannot raise it.
These are weak monotonicity assumptions. A utility may be flat over some range or may assign no value to one outcome dimension. Consequently, a dose that strictly Pareto-dominates another is guaranteed to have at least as high a utility, but equality is possible. If the utility is strictly increasing in every benefit and strictly decreasing in every harm over the compared values, then a strict dominance in at least one dimension guarantees strictly higher utility.
The claim follows by changing one coordinate at a time. Start at \(d_B\)'s outcome vector. Raise each benefit coordinate to its value at \(d_A\), and lower each harm coordinate to its value at \(d_A\). At every step, utility cannot decrease by the assumed monotonicity. After all coordinates have been changed, the vector is \(d_A\)'s, so its utility is at least the starting utility. The ordering of these coordinate changes does not matter to the inequality: each individual change has the required direction while other coordinates are held fixed.
This is an outcome-by-outcome statement. If outcomes vary among patients or scenarios, a corresponding expected-utility guarantee requires a pointwise comparison under a meaningful pairing: for every same patient or scenario being compared, \(d_A\) must provide at least as much of every benefit and no more of every harm. Applying the theorem to each paired outcome gives a utility inequality for each, and averaging those inequalities preserves the ordering when the utilities have finite expectations. A comparison of population averages alone is not that pointwise condition.
Worked Example: Comparing Two Fixed Outcome Profiles
Suppose two hypothetical doses are summarized on the same target-population scale. Benefit 1 is symptom improvement in points, benefit 2 is response probability, harm 1 is toxicity grade on a scale where larger is worse, and harm 2 is serious-adverse-event risk.
| Dose | Symptom improvement (points) | Response probability | Toxicity grade | Serious-event risk |
|---|---|---|---|---|
| \(d_A\) | 6.0 | 0.70 | 1.2 | 0.08 |
| \(d_B\) | 5.5 | 0.65 | 1.5 | 0.11 |
For the two benefits, \(6.0 \geq 5.5\) points and \(0.70 \geq 0.65\). For the two harms, \(1.2 \leq 1.5\) toxicity-grade units and \(0.08 \leq 0.11\) in risk. Every inequality is in \(d_A\)'s favor, so \(d_A\) strictly Pareto-dominates \(d_B\).
For illustration, consider the utility \(U=\ln(1+\text{improvement}/1\text{ point})+2(\text{response probability})-0.1(\text{toxicity grade})^2-3(\text{serious-event risk})\). It is increasing in the two benefits and decreasing in the two harms on these scales. Substitution gives \(U(d_A)=\ln(7)+2(0.70)-0.1(1.2)^2-3(0.08)=1.9459+1.4000-0.1440-0.2400=2.9619\). For \(d_B\), \(U(d_B)=\ln(6.5)+2(0.65)-0.1(1.5)^2-3(0.11)=1.8718+1.3000-0.2250-0.3300=2.6168\). The utility difference is \(2.9619-2.6168=0.3451\) utility units in favor of \(d_A\). The theorem is stronger than this particular calculation: any utility with the stated monotonicity has \(U(d_A)\geq U(d_B)\).
Applying the Result to Variable Outcomes
For individual-level outcomes, the comparison must preserve the pairing that makes “every outcome” meaningful. For example, if the same three hypothetical patient profiles can be evaluated under both doses, compare each patient's benefits and harms across doses. If dose-specific outcomes are uncertain, a pointwise guarantee across the paired profiles implies a guarantee for the average utility, provided the utility expectations exist. This does not require utility to be linear.
Worked Example: Pointwise Dominance and Expected Utility
Consider three hypothetical patient profiles. Benefit is a clinical improvement score, and harm is a toxicity score. For each profile, dose \(d_A\) has at least as much benefit and no more harm than dose \(d_B\). Use \(U(b,h)=\ln(1+b)-0.2h^2\), where the scales are chosen so \(1+b\) is positive.
| Profile | \(b(d_A)\) | \(h(d_A)\) | \(b(d_B)\) | \(h(d_B)\) |
|---|---|---|---|---|
| 1 | 4 | 1 | 3 | 2 |
| 2 | 2 | 0 | 2 | 1 |
| 3 | 6 | 2 | 5 | 3 |
The benefit comparisons are \(4\geq3\), \(2=2\), and \(6\geq5\); the harm comparisons are \(1\leq2\), \(0\leq1\), and \(2\leq3\). Thus dominance holds for every paired profile.
The utilities at \(d_A\) are \(\ln(5)-0.2(1)^2=1.4094\), \(\ln(3)-0.2(0)^2=1.0986\), and \(\ln(7)-0.2(2)^2=1.1459\). Their equally weighted mean is \((1.4094+1.0986+1.1459)/3=1.218\) utility units.
At \(d_B\), they are \(\ln(4)-0.2(2)^2=0.5863\), \(\ln(3)-0.2(1)^2=0.8986\), and \(\ln(6)-0.2(3)^2=-0.0082\). Their mean is \((0.5863+0.8986-0.0082)/3=0.4922\) utility units. The mean utility is greater for \(d_A\), as guaranteed by the profile-by-profile inequalities. This calculation illustrates the result for one utility; the pointwise proof applies to every utility with the required monotonicity.
What Dominance Does Not Establish
A common mistake is to treat better averages as proof that one dose cannot be worse under any monotone utility. A larger mean benefit and a smaller mean harm do not in general show that each patient's benefit is at least as large and each harm no greater. With nonlinear utility, the distribution of outcomes can matter. Conversely, a dose can fail to dominate another even when its mean utility is greater for a particular utility function.
Worked Example: Better Mean Benefit Does Not Guarantee Higher Utility
Suppose harm is zero for both hypothetical doses. Dose \(d_A\) gives benefit \(b=5\) to each of two equally weighted profiles. Dose \(d_B\) gives benefits \(b=0\) and \(b=11\). The mean benefits are \(\bar b(d_A)=5\) and \(\bar b(d_B)=(0+11)/2=5.5\), so \(d_B\) has the larger mean.
For the increasing utility \(U(b)=\ln(1+b)\), dose \(d_A\)'s mean utility is \(\ln(6)=1.7918\) utility units. Dose \(d_B\)'s mean utility is \((\ln(1)+\ln(12))/2=(0+2.4849)/2=1.2425\) utility units. Therefore \(d_B\)'s larger mean benefit does not imply a larger mean utility for this nonlinear preference.
There is no pointwise dominance here: \(d_B\)'s benefit is lower for the first profile and higher for the second. The example does not contradict the theorem; it shows why average outcomes alone cannot replace its componentwise condition.
Diagnostics, Comparisons, and Limits
A practical dominance check begins with the outcomes, not a utility calculation. List every benefit and harm that defines the comparison, confirm the preferred direction of each scale, and compare the two doses dimension by dimension. For patient-level or scenario-level claims, make the pairing explicit. A reversed scale—for example, treating a higher symptom score as better when higher means worse—can reverse the conclusion.
- Check completeness. Omitting a benefit or harm can create an apparent dominance that does not hold for the decision actually being made.
- Check scale direction. State whether larger or smaller values are preferred before comparing numbers.
- Check the comparison basis. Do not describe averages as pointwise comparisons, or combine summaries from different populations without justification.
- Check uncertainty. Estimated outcomes may not establish that the true outcomes satisfy the inequalities. If uncertainty is relevant, describe it rather than treating point estimates as certain.
Pareto dominance differs from weighted-sum utility analysis. A weighted sum can rank doses that trade off benefit against harm, but its ranking depends on the selected weights. Dominance needs no trade-off weight and is robust to every utility in the stated monotone class. A noncompensatory objective or hard safety constraint can impose additional decision rules; those rules must be applied as specified, and they do not change the mathematical meaning of dominance. If a dose is unavailable or infeasible under a constraint, dominance by that dose alone does not make it a usable option.
The limitation is equally important: many candidate doses are incomparable. One may offer more benefit but also more harm, or improve some benefits while worsening others. Pareto dominance cannot settle those trade-offs. Nor does it establish that a dose is clinically worthwhile, safe, or optimal in an absolute sense. It only rules out preferring a dominated option when the comparison is valid and utility is monotone in the declared directions.
Check Your Understanding
Answer each question, then compare your reasoning with the solutions below.
- Dose \(d_A\) has benefit values \(8\) and \(0.40\), and harm values \(2\) and \(0.06\). Dose \(d_B\) has benefit values \(7\) and \(0.40\), and harm values \(2\) and \(0.08\). Does \(d_A\) weakly or strictly dominate \(d_B\)?
- If one dose strictly dominates another, must every weakly monotone utility assign it strictly greater utility? Explain.
- Why does pointwise dominance across paired patient profiles guarantee an ordering of mean utility when the utilities have finite means?
- Two doses have mean benefits of 4 and 5, respectively, and identical mean harms. Does the second dose necessarily have greater expected utility for every increasing utility? Why or why not?
- A proposed utility decreases as a harm score increases, but the score is coded so that larger values mean fewer adverse symptoms. Is that utility monotone in the stated clinical direction?
Solutions. (1) \(d_A\) strictly dominates: its first benefit is higher, its second is equal, its first harm is equal, and its second harm is lower. (2) No. A weakly monotone utility may be flat over the changed outcome range, so it can assign equal utility; strict response to a changed dimension is needed for strict preference. (3) The theorem gives an inequality for each paired profile, and averaging a set of inequalities preserves the direction, provided the utilities have finite means. (4) No. Mean comparisons do not establish pointwise dominance, and a nonlinear increasing utility can rank distributions differently, as in the example above. (5) No. If larger scores mean fewer symptoms, a utility that decreases as the score rises moves in the wrong direction; its monotonicity must be assessed using the clinical meaning of the scale.