Why Define an Outcome Space?
A treatment can affect several clinically important outcomes at once: it may improve a symptom, reduce disease events, and also cause adverse effects. Before comparing those effects or summarizing them, the outcomes must be defined so that each has a clear clinical meaning. An outcome vector makes the distinct attributes visible rather than hiding them in a single number.
The purpose here is to specify what a patient's clinical outcome profile contains. A vector is not itself a judgment that one treatment is preferable, and its coordinates are not automatically interchangeable. In particular, writing several outcomes side by side does not assign them weights or combine them.
For \(J\) attributes, let \(\mathcal{Y}_j\) be the domain of values allowed for attribute \(j\). The Cartesian product \(\mathcal{Y}_1 \times \cdots \times \mathcal{Y}_J\) is an ambient space: it contains every coordinate-wise combination. The set \(\mathcal{F}\) of feasible outcome profiles is a subset of that ambient space. It equals the full product only if every combination is possible under the clinical definitions. For example, if two binary attributes must always agree, only \((0,0)\) and \((1,1)\) are feasible, even though the ambient product also contains \((0,1)\) and \((1,0)\).
Specify Every Coordinate
A coordinate is interpretable only when its measurement is fully stated. At minimum, specify the clinical concept and operational measure, its direction, units or categories, the population to which it applies, and its assessment horizon. These details prevent a vector from mixing measures that sound similar but answer different questions.
- Meaning and measurement: State what is recorded and how. “Disease severity” is not enough; specify, for example, the score and visit used.
- Direction: State whether a higher or lower value represents a clinically better outcome, or whether categories have no simple ordering. This is descriptive orientation, not a preference weight.
- Units and domain: Give units for quantitative values and allowable categories or ranges. For a count, say what is counted and over what period.
- Population: Identify the people described, including relevant eligibility or subgroup restrictions. A measure may be defined only for a clinically specified subgroup.
- Assessment horizon: State the time point or period. A level at week 24 and a change from baseline to week 24 are different attributes.
The profile is often written for one person under a specified treatment regimen, with one coordinate per attribute. If a trial has several regimens, each regimen can have a corresponding set of profiles; defining that notation does not by itself define a treatment-effect estimand. In particular, a population average, a treatment contrast, and an individual's multivariate profile are different objects.
Worked Example: A Change and a Period Count
In a hypothetical study of adults with hypertension, define two attributes for each participant: systolic blood-pressure reduction from baseline to week 24, and the number of cardiovascular events during the 24 weeks. At baseline, one participant has systolic pressure 156 mmHg; at week 24, it is 142 mmHg. The participant has one cardiovascular event during follow-up.
Define the first coordinate as baseline pressure minus week-24 pressure:
A positive value is a reduction, so higher values indicate a larger decrease in systolic pressure. This coordinate is a change, not a week-24 level. Its units are mmHg, and its horizon is baseline to week 24.
The second coordinate is \(Y_2=1\) cardiovascular event during the 24-week period. It is a count with domain \(\{0,1,2,\ldots\}\); lower values indicate fewer events. Thus the participant's profile is \((14\ \text{mmHg},\,1\ \text{event})\). Its coordinates retain different units and clinical meanings; the vector notation does not turn them into a common scale.
Domains and Feasible Profiles
The domain of each coordinate describes values that are meaningful for that attribute. A binary event indicator might have domain \(\{0,1\}\), with 1 defined as event observed by the horizon. A count might have nonnegative integer values. A continuous measure might have a clinically or instrument-defined range. Domain definitions should not imply that all combinations across attributes are attainable.
To define the outcome space, first form the ambient product of the coordinate domains, then impose any known structural constraints. For instance, if an event can occur only after a participant enters a risk period, the event measure's definition must say how participants who never enter that period are represented. If two attributes are logically linked by their definitions, their joint feasible set must reflect that link. A constraint should be based on the clinical definitions, not merely on the fact that a combination was absent from a small sample.
Worked Example: Binary Attributes With a Structural Constraint
Suppose a hypothetical condition has two recorded indicators at a specified visit: \(Y_1=1\) if the condition is present and \(Y_2=1\) if the same condition is recorded as active. For this example, the protocol defines these indicators identically for every participant, so they must agree. Each coordinate has domain \(\{0,1\}\).
The ambient product has four profiles:
But the definition requires \(Y_1=Y_2\), so the feasible set is
The profiles \((0,1)\) and \((1,0)\) are in the ambient product but violate the stated constraint. This distinction matters: saying that every binary combination is possible would incorrectly describe this outcome space. In an actual application, two attributes should be defined separately only if their clinical or measurement meanings justify recording them separately.
Worked Construction for a Trial
A useful test of a proposed vector is whether an analyst could assign each participant's coordinates from the protocol and clinical data without making unstated choices. The next example shows how a time limit and event definition belong in the attribute itself.
Worked Example: Building a Three-Attribute Profile
Consider a hypothetical trial in adults with a chronic inflammatory disease, assessed over 16 weeks. Define: (1) change in a symptom score from baseline to week 16, where lower scores are better; (2) number of disease flares during weeks 0–16, where fewer is better; and (3) whether a serious adverse event occurred during weeks 0–16, coded 1 for yes and 0 for no. For illustration, the score ranges from 0 to 10.
A participant has a symptom score of 6 at baseline and 4 at week 16, has two flares, and has no serious adverse event. Define the score-change coordinate as week-16 score minus baseline score:
Because lower symptom scores are better, a negative change represents improvement. The remaining coordinates are \(Y_2=2\) flares and \(Y_3=0\) (no serious adverse event). The profile is \((-2\ \text{score points},\,2\ \text{flares},\,0)\). Its coordinates have distinct domains and units: the score change is bounded by the score's possible baseline and follow-up values, the flare count is a nonnegative integer, and the adverse-event indicator is binary.
This specification also makes the horizon explicit: the score is measured at week 16 relative to baseline, while both event attributes count events during weeks 0–16. It does not state how the three coordinates should be traded off against one another; that is a separate decision question.
Justification: Why Keep the Coordinates Explicit?
The vector representation preserves the separate values before any summary is considered. Formally, if a profile is \(y=(y_1,\ldots,y_J)\), its \(j\)th coordinate is obtained by projection:
Thus each specified clinical measure can be inspected in its own units and direction. A single scalar summary, by contrast, would require an explicit rule that maps profiles to numbers. If that rule has not been defined, no calculation can legitimately replace the vector with a scalar. This is why the vector is a useful starting representation even when a later analysis will compare or summarize treatment outcomes.
Name measures that represent distinct, clinically relevant outcomes; do not use vague labels in place of operational definitions.
For each attribute, state its measure, direction, units or categories, population, and visit or period.
List coordinate-wise domains, then identify structural restrictions on which combinations can occur.
Verify that each participant's values can be assigned consistently, including the handling of events, eligibility, and unavailable measurements.
Diagnostics, Alternatives, and Limitations
Several warning signs indicate that an outcome space is not yet well defined. A coordinate named only “response” lacks a threshold or classification rule. A “change” coordinate without an order of subtraction has ambiguous direction. A count without a period is incomplete. A measure without units or a population can be misread, and a profile with mismatched assessment horizons may juxtapose outcomes that refer to different periods without making that fact clear.
Missing or unobserved values also need careful treatment. Missingness is not automatically a clinical outcome value: it should not be coded as zero, no event, or a poor outcome unless that coding is part of a justified and explicitly stated definition. Likewise, for an event measure, the specification should distinguish “no event by the end of follow-up” from “event status unknown.” These distinctions affect which profile can be assigned; they are not resolved by the vector notation.
The main alternatives answer different questions. A single endpoint focuses on one measure, which can be clear and operationally simple but may omit other important outcomes. A composite endpoint combines components according to a defined event or scoring rule; that may simplify a primary analysis, but the resulting value can conceal which components differ. A vector keeps attributes separate and makes their definitions visible, but by itself it does not rank profiles or provide one summary for a decision. No representation removes the need to justify the clinical measures, population, and time horizon.
A well-specified outcome space is a practical foundation for later benefit-risk work because it makes clear what is being described before any comparison or aggregation is attempted. Its quality can be checked with three questions: Is every coordinate clinically interpretable? Can its value be assigned using the stated measurement rule and horizon? Are the feasible joint profiles consistent with the clinical definitions? If any answer is uncertain, refine the definitions before proceeding.
Check Your Understanding
Answer each question, then compare with the solutions below.
- A study records a symptom score at baseline and week 12. What additional definition is needed before “symptom change” has an unambiguous sign?
- Two binary attributes each have domain \(\{0,1\}\), but the protocol requires them to agree. What are the ambient product and feasible set?
- A patient has a baseline score of 8 and a week-12 score of 5. If the coordinate is defined as week-12 minus baseline, calculate the change and interpret its direction when lower scores are better.
- Why should an unrecorded adverse-event status not automatically be entered as 0?
Solutions. (1) State the subtraction order, such as week-12 minus baseline, as well as the score's units and assessment horizon. (2) The ambient product is \(\{(0,0),(0,1),(1,0),(1,1)\}\); the feasible set is \(\{(0,0),(1,1)\}\), because agreement requires equal coordinates. (3) \(5-8=-3\) score points. The negative change means the score decreased by 3 points, which is improvement under the stated direction. (4) An unrecorded status is unknown, not necessarily “no event”; coding it as 0 would assert an outcome not established by the data.