Why Read a Pareto Front and a Sensitivity Profile?
Regimen optimization often involves competing goals: greater efficacy may come with greater toxicity. A single “best” regimen can conceal how much the recommendation depends on the tradeoff chosen between those goals. A Pareto front shows which regimens remain competitive when neither objective is reduced to a single score. A sensitivity profile then shows how the preferred regimen changes as a specified decision weight or model assumption changes.
These displays answer different questions. The front describes tradeoffs among candidate regimens; the profile describes how a decision rule ranks those candidates across a range of inputs. Neither display, by itself, establishes that a regimen is clinically acceptable. The objective definitions, constraints, and plausible input ranges must be justified in context.
Dominance depends on the direction and definition of each objective. If toxicity is instead expressed as a safety score to maximize, the inequality must be changed accordingly. The front also depends on the candidate set: excluding a regimen can change which remaining regimens appear efficient. Pareto efficiency is not a claim that every regimen on the front is acceptable or equally desirable.
Reading the Front Before Applying Weights
Consider hypothetical efficacy scores and toxicity-burden scores, both expressed in points on scales defined for this example. Higher efficacy is better and lower toxicity burden is better. Regimens A, B, and C each offer a tradeoff: one has more efficacy but also more toxicity burden than another. None dominates another, so all three lie on the front. Regimen D, by contrast, is worse than A on both objectives and can be removed from consideration under any decision rule that prefers higher efficacy and lower toxicity.
Worked Example: Finding Dominated Regimens
A hypothetical model gives the following objective values. Treat the scores as decision-model outputs, not observations from a real study.
| Regimen | Efficacy \(E_i\) (points) | Toxicity burden \(T_i\) (points) |
|---|---|---|
| A | 70 | 30 |
| B | 82 | 48 |
| C | 62 | 20 |
| D | 65 | 35 |
A has greater efficacy and lower toxicity burden than D: \(70>65\) and \(30<35\). Thus A dominates D, and D is not on the Pareto front. A does not dominate C because A has greater efficacy but also greater toxicity burden. Similarly, B has greater efficacy but greater toxicity burden than A, and C has lower toxicity burden but lower efficacy than A. These comparisons show that A, B, and C are nondominated in this candidate set.
The conclusion is limited to the stated objectives and candidates. If an additional objective such as dosing convenience is included, dominance must be reassessed using that objective too. If a safety constraint excludes some candidates, the relevant front is the front among candidates satisfying that constraint.
From a Front to a Weight-Specific Ranking
A common way to select among Pareto-efficient regimens is a weighted utility. For this tutorial, let \(w\) be the relative weight on efficacy, with \(0\le w\le1\); the remaining weight, \(1-w\), is assigned to toxicity burden. The illustrative utility is \(wE_i-(1-w)T_i\). It is a linear scoring rule, not a universal clinical utility model. Its use assumes that the objectives have been placed on scales for which this arithmetic and the chosen weights have a defensible interpretation.
For two regimens, equating their utilities gives a direct derivation of the switching weight. If regimen \(i\) has greater efficacy but greater toxicity burden than regimen \(j\), the tie occurs at:
Rearranging yields \(w(E_i-E_j)=(1-w)(T_i-T_j)\). When both differences are positive, the solution is \(w=(T_i-T_j)/[(E_i-E_j)+(T_i-T_j)]\). This expression only identifies the pairwise tie. To determine the preferred regimen overall, compare all candidates on each interval separated by the pairwise switching points.
Worked Example: Exact Switching Weights and Rankings
Use A \((E_A=70,\ T_A=30)\), B \((E_B=82,\ T_B=48)\), and C \((E_C=62,\ T_C=20)\). Their utilities are:
For A and C, set \(100w-30=82w-20\). Then \(18w=10\), so the exact tie is \(w=10/18=5/9\), approximately \(0.5556\). For A and B, set \(100w-30=130w-48\). Then \(30w=18\), so the exact tie is \(w=0.6\).
To check which regimen is best between the ties, take \(w=0.58\). Then \(U_A=100(0.58)-30=28\), \(U_B=130(0.58)-48=27.4\), and \(U_C=82(0.58)-20=27.56\); A is preferred. At \(w=0.5\), \(U_A=20\), \(U_B=17\), and \(U_C=21\), so C is preferred. At \(w=0.8\), \(U_A=50\), \(U_B=56\), and \(U_C=45.6\), so B is preferred.
The complete ranking is therefore C for \(0\le w<5/9\), A for \(5/9<w<0.6\), and B for \(0.6<w\le1\). At \(w=5/9\), A and C tie for the maximum; at \(w=0.6\), A and B tie for the maximum. The exact fraction matters: \(0.5556\) is a rounded display of \(5/9\), not the boundary itself. At \(w=0.5556\), A is slightly preferred to C, not tied.
Sensitivity Profiles: What Changes, and Why?
A sensitivity profile plots, tabulates, or otherwise reports the preferred regimen as one or more inputs vary. A profile over \(w\) is a weight sensitivity profile: it shows which regimen is best for each efficacy-toxicity tradeoff under a fixed set of model outputs. A profile over alternative parameter sets or structural assumptions is a model-assumption sensitivity profile: it shows whether the recommendation changes when the inputs used to predict efficacy and toxicity change.
Keep these axes distinct when interpreting a profile. Changing \(w\) represents a change in the decision maker’s relative valuation of the objectives, not a change in the exposure-response model. Changing model assumptions represents uncertainty or alternative plausible representations of outcomes, not a change in the stated tradeoff. A useful report identifies what varied, what was held fixed, and whether the displayed result is a ranking, a utility value, or an outcome prediction.
Worked Example: Comparing Assumption-Specific Profiles
Suppose the same three hypothetical regimens are evaluated under two plausible model-assumption sets. In the first set, their objective values are those above. In a second set, a changed toxicity-response assumption gives A \((70,34)\), B \((82,48)\), and C \((62,20)\). For each set, retain the same weighted-utility definition and vary \(w\).
In the second set, \(U_A(w)=70w-34(1-w)=104w-34\); \(U_B(w)=130w-48\); and \(U_C(w)=82w-20\). The A–C tie solves \(104w-34=82w-20\), giving \(22w=14\), or \(w=7/11\approx0.6364\). The A–B tie solves \(104w-34=130w-48\), giving \(26w=14\), or \(w=7/13\approx0.5385\).
Here the pairwise tie order differs from the first set: the A–B tie occurs before the A–C tie as \(w\) increases. Check the overall maximum rather than assuming every pairwise tie is a change in the winner. At \(w=0.58\), the second-set utilities are \(U_A=26.32\), \(U_B=27.4\), and \(U_C=27.56\), so C is preferred. At \(w=0.65\), they are \(33.6\), \(36.5\), and \(33.3\), so B is preferred. At \(w=7/13\), A and B each have utility \(104(7/13)-34=130(7/13)-48=22\), while C has utility \(82(7/13)-20=314/13\approx24.1538\), so C is preferred. The A–B tie is not a switch in the overall winner. B and C tie at \(w=7/12\); C is preferred for \(0\le w<7/12\), and B for \(7/12<w\le1\). A is never preferred overall in the second set. The A–C crossing at \(7/11\) does not displace B from the maximum.
This example illustrates why pairwise boundaries alone do not define a profile: every candidate’s utility must be compared at boundaries and between them. Comparing the two assumption-specific profiles also reveals a decision that is sensitive to assumptions in some weight regions, even though it does not identify which assumption is correct.
Diagnostics, Comparisons, and Limitations
A profile is most informative when its construction is auditable. Check the following before interpreting apparent switches:
- Objective direction and units: verify that efficacy is consistently treated as beneficial and toxicity burden as harmful, and that both scores have documented scales.
- Candidate completeness: an omitted feasible regimen may dominate a displayed one or become preferred under some weights.
- Boundary arithmetic: calculate switching points from unrounded values, then label any displayed decimal as approximate. Check all candidates at each boundary and in each interval.
- Assumption isolation: state whether each profile changes decision weights, model parameters, structural choices, constraints, or several at once.
- Near ties: a very small utility difference can make a nominal winner fragile. Report the size of the difference or show the competing candidates rather than presenting the winner as unequivocal.
The Pareto front is a useful alternative to a single weighted score when decision makers do not wish to specify weights yet. It narrows the set to nondominated candidates, but does not select one. Weighted utility gives a ranking conditional on a scoring rule and weights, but compresses multiple objectives into one number. Constraint-based selection is another alternative: require, for example, toxicity burden below a prespecified limit, then compare efficacy among regimens that meet it. A constraint can be easier to interpret when a threshold is clinically meaningful, but the threshold itself must be justified, and the result can change sharply near it.
Linear weighted utility has important limitations. It assumes the scores are commensurate in the specified way and that tradeoffs are constant across the range: the same change in efficacy has the same weighted value regardless of the starting level. It can also produce misleading rankings if the scales or weights are arbitrary. Nonlinear utility or additional constraints may be needed when preferences have thresholds, diminishing returns, or unacceptable-risk regions. Sensitivity analysis exposes dependence on chosen inputs; it does not eliminate that dependence, assign probabilities to assumptions, or demonstrate that a model is clinically valid.
Common Pitfalls
- Treating every Pareto-efficient regimen as equally good. Efficiency means only that no candidate is at least as good on every objective and strictly better on one. It does not make the tradeoffs equally acceptable.
- Calling a rounded decimal the exact switching point. For example, \(0.5556\) is an approximation to \(5/9\). Use the exact value for tie claims and interval definitions; use the decimal only as a clearly labelled approximation.
- Assuming a pairwise tie must change the overall winner. A third regimen may have higher utility at the crossing. Verify the maximum across all candidates.
- Describing a change in weights as model uncertainty. Weights encode preferences; model assumptions determine predicted objective values. Report the two sensitivities separately.
- Equating a high utility with clinical acceptability. A weighted score cannot override a safety constraint unless that constraint is explicitly incorporated into the decision rule.
Check Your Understanding
Answer each question, then compare your reasoning with the solutions below.
- Regimen X has efficacy 50 and toxicity burden 12; regimen Y has efficacy 48 and toxicity burden 10. Does either dominate the other? Explain.
- For two regimens with \((E_1,T_1)=(60,25)\) and \((E_2,T_2)=(50,15)\), calculate the exact tie weight using \(U_i(w)=wE_i-(1-w)T_i\).
- Why should all candidate utilities be checked at a pairwise switching weight?
- A profile changes only the efficacy-toxicity weight \(w\), while model-predicted outcomes remain fixed. Is this model-assumption sensitivity or weight sensitivity? What does it vary?
Solutions. 1. Neither dominates. X has higher efficacy, but also higher toxicity burden; Y has lower toxicity burden, but lower efficacy. 2. Set \(60w-25(1-w)=50w-15(1-w)\). This gives \(85w-25=65w-15\), so \(20w=10\) and \(w=1/2\). At that weight both utilities equal \(17.5\) points. 3. A third regimen may have greater utility than both tied regimens, so their tie may not be a tie for the overall maximum or a change in the preferred regimen. 4. It is weight sensitivity. It varies the decision maker’s relative valuation of efficacy versus toxicity while holding the model’s predicted objective values fixed.