Pareto Frontier

Pareto Frontier

A Pareto frontier is the set of feasible choices for which no other feasible choice improves at least one stated objective without worsening another. It helps decision makers see trade-offs when no single option is best on every dimension. The frontier depends on which options are feasible, which objectives are included, how each is measured, and whether larger or smaller values are preferred.

Dominance and the frontier

Write each feasible option $x$ as an objective vector $g(x)=(g_1(x),\ldots,g_k(x))$, orienting all components so that larger is better. Option $a$ Pareto-dominates $b$ if $g_j(a)\geq g_j(b)$ for every objective $j$ and $g_j(a)>g_j(b)$ for at least one objective. Thus a dominating option is at least as good in all stated dimensions and strictly better in one or more. The Pareto frontier of a feasible set $F$ is

$$P=\{x\in F:\nexists y\in F\text{ that Pareto-dominates }x\}.$$

Costs can be represented as their negatives when using a “larger is better” vector, or left as a “smaller is better” objective with the inequalities reversed for that component. Two options with identical objective vectors do not strictly dominate each other; both may remain in the set of efficient options, even though they occupy one point in objective space. An option on the frontier is also called non-dominated or Pareto-efficient relative to the defined feasible set and objectives.

This is a filter, not a complete ranking. Two non-dominated options can trade better health against higher cost, or aggregate health against an equity outcome; choosing between them requires preferences, constraints or another decision rule. A poor score on an omitted outcome does not enter the dominance test, which is why careful objective selection is essential.

Worked health allocation example

Suppose a programme planner has five mutually exclusive, feasible service packages. Total programme cost is to be minimised and total QALYs gained in the same population and time horizon are to be maximised. These values are invented solely for teaching, with costs in thousands of currency units.

Here $P=\{A,B,D\}$ among these five packages. B strictly dominates C because $120<130$ and $8>7$; D strictly dominates E because $150<160$ with equal QALYs. A, B and D each retain a trade-off: moving from A to B or B to D improves health but increases cost. Removing a dominated option is justified only for the stated objectives and feasible alternatives; if C alone had a separately valued access or equity advantage, that dimension would need to be represented and its status reassessed.

The QALY gain must mean the same thing for every package. If one number counts lifetime gain and another counts one-year gain, or costs use different perspectives, the apparent frontier is not a sound comparison. In practice, uncertainty intervals, model assumptions and resource interactions must also be examined before calling an option robustly dominated.

From frontier to a decision

If a budget cap is binding, examine the feasible packages under that cap; a package outside the cap is not a selectable alternative even if it is non-dominated in an unconstrained plot. If the objective is to maximise $\lambda Q-C$ at a stated willingness-to-pay threshold $\lambda$, calculate net monetary benefit for each feasible package and choose the maximum under the stated rule. A point's presence on the Pareto frontier alone does not mean it maximises net benefit for any specified threshold.

Using the example in thousands of currency units and $\lambda=10$ thousand per QALY gives net benefits $10(5)-100=-50$ for A, $10(8)-120=-40$ for B, and $10(10)-150=-50$ for D. B has the highest net benefit of these three at that threshold. These totals are relative to the common origin used in the illustrative table; adding the same comparator's costs and effects to every alternative does not change their pairwise ranking. The numerical threshold is illustrative and does not assert any real-world policy threshold.

With more than two objectives, a decision maker might consider patient outcomes, unmet need, distributional effects, implementation capacity and budget impact as separate dimensions. The resulting frontier can be a surface rather than a line, and some objectives may be hard constraints rather than tradeable scores. Weights or an explicit utility function can then select an option, provided the underlying scales, trade-offs and stakeholder judgments are clear.

Relationship to cost-effectiveness frontiers

A two-dimensional cost–effectiveness plot with lower cost and greater effect preferred uses the same basic dominance test. Conventional cost-effectiveness analysis may additionally remove an extendedly dominated strategy when a combination or sequence of other strategies provides better value under the relevant decision framework. This stronger criterion depends on assumptions about divisibility, mixtures and the decision rule; it is not identical to basic Pareto dominance.

Similarly, a weighted-sum optimisation can find some efficient solutions but may miss non-convex portions of a frontier, particularly when programmes are indivisible. Do not draw a continuous frontier between discrete alternatives unless mixtures, fractional allocation or a modelled continuum are genuinely feasible. If an alternative appears attractive only under an impossible fractional package, it is not an available choice.

Constructing and checking a frontier

First define the perspective, population, time horizon, candidate options, common units and direction of preference for each objective. Then identify feasibility conditions, including budgets, staffing, capacity and mutually exclusive selections. Calculate objective values consistently, compare each option with every feasible competitor, and retain those not dominated. For a large continuous choice set, use a suitable multiobjective optimisation algorithm and assess coverage of the frontier rather than treating one computed solution as exhaustive.

Check whether any removed option truly has a feasible competitor that is no worse on all included objectives and strictly better on at least one. Record the comparator that dominates it, and preserve a tie when all measured objectives coincide. Recompute as objectives, constraints, estimates or scenarios change; the frontier is conditional on these choices. Small numerical differences should be assessed against measurement accuracy rather than silently interpreted as real improvements.

Uncertainty and interpretation

The simple test assumes known objective values. Health economic costs and effects are estimated, so an option that looks dominated at point estimates may not be dominated across plausible parameter values. Repeat the comparison in probabilistic draws or decision-relevant scenarios and describe which strategies are on the frontier under each; the probability of being non-dominated is not the probability of being the best choice at a specified set of preferences.

One frontier does not certify an ethically acceptable or affordable policy. An option may be non-dominated because it serves a valued objective while still violating a minimum standard, and an efficient point may distribute benefits inequitably if distribution is not represented. Discuss the opportunity costs of resources, fairness and implementation alongside frontier membership, and keep causal or modelling uncertainties visible.

The ISPOR MCDA task force introduction describes the role of multiple objectives in healthcare decisions; its emerging good practices report addresses implementation and constraints. These sources guide the decision context. The mathematical dominance rule above is stated explicitly so that frontier membership in a particular example can be independently checked.