Anchoring Utility Scales to Clinical Consequences

Why Clinical Anchors Matter

A utility scale is useful only if its numbers have a defensible connection to consequences that matter to patients. A score of 70 might look precise, but without defined reference states and an elicitation rule, it does not tell a decision-maker what clinical experience earns that score or whether a gap of 10 points matters. Anchoring makes the scale interpretable: it identifies the clinical outcomes represented by the ends of the scale and gives intermediate values a stated preference meaning.

The task is not to make a scale universally correct. It is to define a scale for a specified decision context, then check whether its numbers agree with preferences elicited in that context. A score difference is not automatically a clinically important difference. It becomes interpretable when the reference states, elicitation procedure, and criterion for meaningfulness are made explicit.

Before eliciting values, specify the population, time horizon, treatment context, and outcome descriptions. For example, “grade 2 fatigue” is incomplete if its duration is unknown and the scale is meant to compare experiences over six months. Anchors should describe profiles rather than vague labels, and they should be plausible enough that respondents can understand the consequences being compared.

A common convention assigns 0 to a defined least-preferred reference profile and 100 to a defined most-preferred profile. These are local anchors, not claims that the worst imaginable health state has value 0 or that the best imaginable health state has value 100. If either reference profile changes, the interpretation of the numbers must be reconsidered. The reference states should be held fixed while values on that scale are elicited and compared.

Assigning Values Through Preference Indifference

One way to attach preference meaning to a number is to ask whether a person would prefer a clinical state or a lottery between the two anchors. Suppose the best reference state is \(B\), the worst reference state is \(W\), and an intermediate state is \(x\). Ask for the probability \(p\) of receiving \(B\) in a hypothetical lottery, with probability \(1-p\) of receiving \(W\), that makes the respondent indifferent between the lottery and state \(x\).

The following assignment requires a specific assumption: the respondent can understand and compare the described outcomes and the hypothetical lottery, and preferences over the lottery are represented by its probability-weighted utility. It also assumes that the same anchors and context are used for all states being compared. If indifference is unstable or the lottery is not credible to respondents, the resulting numbers should not be treated as precise preference measurements.

The assignment follows directly from the indifference condition. Under probability-weighted utility, the lottery’s utility is the probability-weighted average of its two anchor utilities. Equating that quantity to the utility of \(x\) gives:

Thus, if two states have indifference probabilities \(p_x\) and \(p_y\), their difference on this scale is:

This relation gives the numerical gap a specific interpretation: it is 100 times the difference in the probabilities of the best anchor in the respective indifference lotteries. A 12-point gap corresponds to a 0.12 difference in those probabilities, given the same anchors and elicitation context. It does not mean that one state is 12% more likely, nor does it by itself establish that patients consider the gap important.

Testing Whether Score Differences Have Meaning

The algebra gives a way to interpret a difference, but not a universal threshold for importance. That threshold must come from the decision context and the relevant stakeholders. A team may ask respondents what change in anchor-lottery probability would be large enough to alter a choice, or it may prespecify a minimum preference difference for a particular decision. The criterion should be stated before using the scores to compare options.

The earlier material on swing weighting calibrates the relative importance of changes in attributes across specified swings. Clinical anchoring adds explicit outcome profiles and a reference-based interpretation for the resulting scale. Neither step alone guarantees that the final numbers capture all relevant preferences. In particular, a ranking tells which state is preferred, but it does not establish how much more it is preferred.

A useful check is to elicit the same preference in two forms. Respondents can provide a direct rating for a state, then answer an indifference-lottery question for that state. Large disagreement may indicate unclear outcome descriptions, difficulty with probabilities, or an inconsistent response. This is a diagnostic, not a rule that one answer must always be correct. If direct ratings are retained, their interpretation should be described separately from lottery-calibrated scores.

Assumptions, Diagnostics, and Alternatives

Interpretation depends on both the clinical descriptions and the elicitation process. The principal checks are practical as well as mathematical:

• Anchor clarity: respondents can distinguish the best and worst profiles, and each includes the same relevant time horizon and outcome context.
• Comprehension: respondents understand the lottery probabilities and can explain the comparison in their own words.
• Consistency: repeated or related questions do not produce unexplained reversals or large contradictions.
• Relevance: the profiles span consequences relevant to the decision, rather than convenient but clinically uninformative extremes.
• Meaningfulness: any threshold for a consequential difference is elicited or justified for the decision; it is not assumed to be a universal number of utility points.

If respondents struggle with the lottery, a direct rating scale may be easier to administer, but its numerical spacing may not represent equal preference differences. Pairwise ranking is simpler still, but yields order rather than magnitude. A time trade-off question can anchor preferences through choices involving duration, but it changes the elicitation task and may be unsuitable when survival duration is not relevant or when trading time is unacceptable. These approaches are alternatives for eliciting preferences, not interchangeable interpretations of the same numbers.

A failure mode is to select the most extreme observed outcomes as anchors without checking whether they are credible or relevant. This can make intermediate values depend on rare or implausible profiles. Another is to let different respondents use different interpretations of an anchor; scores then do not share a stable clinical reference. Record the anchor wording and elicitation procedure alongside the values, and revisit them if the population, time horizon, or decision changes.

Anchoring cannot remove uncertainty in preferences, resolve disagreement among stakeholders, or correct incomplete outcome descriptions. Lottery responses may be sensitive to risk attitudes, probability presentation, or the hypothetical nature of the question. A single score can conceal meaningful variation across patients. When those limitations matter, preserve respondent-level results or report sensitivity to plausible values rather than treating an average as definitive.