From Separate Outcomes to a Joint Preference Value
A dose can improve efficacy while also causing harm. Reporting those outcomes separately is essential, but it does not by itself express how a decision-maker compares a particular combination of benefit and risk. A joint utility function makes that comparison explicit by mapping each patient’s efficacy and safety outcomes to a single preference value.
The utility is not a new clinical endpoint and it does not replace the underlying outcomes. It is a specification of preference over those outcomes. To make it interpretable, define what is measured, for whom, and over what period before assigning any scores. The earlier material on therapeutic window objectives considers efficacy and toxicity thresholds; a joint utility goes further by assigning preference values to combinations of outcomes rather than only classifying whether each threshold is met.
“Patient-level” matters: the function assigns a value to one patient’s joint outcome, not directly to a dose or a population. A dose may be associated with a distribution of such outcomes, but choosing how to summarize that distribution is a separate decision. Here the task is to define the outcome states and their preference values clearly enough that the mapping can be applied consistently.
Specify the Outcomes Before the Scores
A utility table is only as clear as its row and column definitions. For each outcome, state the measurement, direction of improvement, time window, and the rule that places an observation into a category. For safety, also clarify whether the scale reflects the worst event, a particular event type, duration, reversibility, or some other clinically relevant feature. Otherwise, two people can have the same label but quite different experiences.
The following hypothetical specification uses a symptom score from 0 to 10, where a decrease indicates improvement, and records the change from baseline at 12 weeks. It groups changes into three clinically defined efficacy states. Safety is classified over the same 12-week period according to the most serious treatment-emergent event. These categories are examples; actual cut points and event definitions require clinical justification for the condition and treatment being studied.
Once categories have been defined, a joint table provides a direct way to state preferences for every combination. Each cell must refer to a possible joint outcome state. If a combination is impossible by definition, say so rather than silently assigning it a score. If relevant information is omitted, such as the duration of a serious event, state that limitation: the table cannot distinguish patients whose recorded categories are identical.
Worked Example: Building a Joint Preference Table
Suppose a hypothetical clinical team defines efficacy at week 12 by symptom-score reduction: less than 2 points is “no meaningful improvement,” 2 to less than 5 points is “meaningful improvement,” and 5 or more points is “substantial improvement.” Safety is the most serious treatment-emergent event during the 12 weeks: none or mild, moderate and reversible, or serious or irreversible. The team uses a 0-to-10 preference scale, where 0 is the least preferred state in the table and 10 is the most preferred. It judges the reference state of no meaningful improvement with no event to be worth 5 points.
A hypothetical elicitation produces the following scores. They are illustrative preference judgments, not estimates derived from clinical data.
| Efficacy state | None or mild event | Moderate, reversible event | Serious or irreversible event |
|---|---|---|---|
| No meaningful improvement | 5 | 3 | 0 |
| Meaningful improvement | 7 | 5 | 1 |
| Substantial improvement | 10 | 8 | 2 |
For a patient with a 3-point symptom reduction and a moderate, reversible event, the observed state is “meaningful improvement” and “moderate, reversible event.” By the table, \(U_i=u(E_i,S_i)=5\) points. The assigned value records the team’s stated preference for that joint state; it is not a probability, a symptom-score change, or a measure of the event’s frequency.
The table makes the clinical interpretation inspectable: at each efficacy level, a more serious safety category has a lower score; within each safety category, greater efficacy has a higher score. Whether the size of a score difference also has meaning is a further assumption, discussed below.
Interpreting and Using the Score Scale
A utility score needs anchors. Without them, a value such as 5 has no clinical interpretation beyond its position relative to other entries. In the example, the team identifies the best and worst states represented and names a reference state. If scores are rescaled for convenience, the mapping should be stated and should not change the intended preferences.
For instance, a team may convert the 0-to-10 scores in the table to a 0-to-100 point scale by multiplying by 10. If a patient’s table score is 8, then \(U_{100}=10(8)=80\) points. The result is an 80-point score on that rescaled preference scale, not an 80% chance of benefit or a statement that the patient achieved 80% of a clinical response.
A utility scale can express only an ordering, or it can make stronger claims about the sizes of differences. If the scores indicate only that one state is preferred to another, the ordering is informative but a difference of 2 points need not represent the same preference change everywhere on the scale. Treating differences as quantitatively comparable requires a defensible cardinal interpretation of the elicited values. A numerical format alone does not establish that interpretation.
Worked Example: Applying and Rescaling the Table
A hypothetical patient has a symptom-score reduction of 6 points at week 12 and experiences a moderate, reversible event during follow-up. Under the stated definitions, a 6-point reduction is “substantial improvement,” and the safety state is “moderate, reversible event.”
The joint mapping is \(U_i=u(E_i,S_i)=8\) points on the 0-to-10 scale. On the rescaled scale, \(U_{100}=10U_i=10(8)=80\) points. A second hypothetical patient has no meaningful improvement and a serious or irreversible event, giving \(U_i=0\) and \(U_{100}=10(0)=0\) points.
These calculations apply the stated mapping; they do not estimate a treatment effect or establish that one dose is preferable. The clinical interpretation is conditional on the outcome definitions and preference judgments used to build the table.
Check the Preference Structure
Before applying a utility function, inspect it for contradictions and gaps. A useful first check is directional coherence: holding safety fixed, an efficacy improvement that is defined as better should not lower utility; holding efficacy fixed, a worsening safety state should not raise utility. These conditions do not determine the exact scores, but they can flag a table that conflicts with its own stated preferences.
This check is not a substitute for clinical judgment. Some outcome combinations may raise genuine tradeoffs, and there may be no universally correct score. The relevant question is whether the specified values reflect the intended decision context and whether important distinctions have been retained in the recorded outcomes.
Worked Example: Finding and Repairing an Incoherent Entry
Consider a candidate table with three efficacy states and two safety states. In the “none or mild event” column, the scores are 5, 7, and 10 as efficacy improves. In the “serious or irreversible event” column, they are 0, 1, and 2. Suppose the proposed scores in the “moderate, reversible event” column are 3, 5, and 4, respectively.
The last entry violates the stated direction for efficacy: moving from meaningful to substantial improvement while holding the safety state fixed changes the score from 5 to 4. That is a decrease despite an efficacy improvement. The table therefore fails the directional coherence check.
A coherent replacement must be at least 5 in the substantial-improvement, moderate-event cell, because the meaningful-improvement cell is 5. It should also be no greater than 10 if the best state is anchored at 10. For illustration, if the preference judgment supports a score of 7.5, then the rescaled value is \(U_{100}=10(7.5)=75\) points. The inequality \(5\leq 7.5\leq 10\) satisfies this particular check. Coherence narrows the acceptable values; it does not prove that 7.5 is the clinically correct value.
Diagnostics, Alternatives, and Limitations
A practical review should ask whether another analyst could assign a score from the recorded data without guessing. If not, clarify category boundaries, event handling, or the time window. Also check whether the table distinguishes outcomes that matter clinically, whether all relevant combinations are represented, and whether the score ordering matches the stated direction of benefit and harm.
A joint table is transparent when the number of outcome categories is manageable: each score can be inspected directly. Its cost is that the number of cells grows as outcomes and categories are added. A list of separate efficacy and safety summaries keeps dimensions visible, but does not specify a single preference for each joint state. A single composite score is more compact, yet can obscure which outcomes produced the value. The appropriate representation depends on the decision question and on how much detail must remain visible.
A table also compresses information. If efficacy categories omit duration of response, or safety categories omit event reversibility, patients with different experiences may receive the same utility. If follow-up differs between patients, the same labels may not represent comparable outcomes. Missing outcomes, death, treatment discontinuation, and competing events require explicit handling; they should not be assigned an improvised score after seeing the data.
Finally, numerical preferences depend on context. A score set for one disease, treatment duration, or decision-maker may not transfer to another. Stakeholders may disagree, and a sensitivity analysis can reveal whether conclusions depend on disputed entries. The utility function should therefore be documented as a clinical preference specification, including who supplied the judgments and what assumptions they made, rather than presented as an objective property of the drug.
Check Your Understanding
Answer each question using the definitions and checks in this tutorial.
- A hypothetical table assigns 6 points to meaningful improvement with no event and 8 points to substantial improvement with no event. What preference ordering does this express for those two states?
- On a 0-to-10 scale, a patient’s joint outcome receives 7 points. What is the score on a 0-to-100 scale if the specified transformation is multiplication by 10? What does that score not mean?
- A table gives a lower score to substantial than meaningful improvement when the safety state is held fixed. Which check does this fail?
- Why must the follow-up period be specified for both efficacy and safety outcomes?
- If a score table gives only an ordering of states, can a 2-point difference automatically be interpreted as a larger preference gain than a 1-point difference elsewhere on the scale?
Solutions. (1) It expresses a preference for substantial over meaningful improvement in the no-event state, because \(8>6\). (2) The transformed score is \(10(7)=70\) points. It is not a 70% probability of benefit or a percentage response. (3) It fails directional coherence for efficacy: holding safety fixed, improvement should not reduce utility. (4) The time window determines which outcomes count; different windows can change both the recorded state and its clinical meaning. (5) No. An ordering alone does not establish that score differences are comparable in size; that requires a defensible cardinal interpretation.