Why Weight the Full Swing?
In a multi-attribute benefit-risk model, an attribute weight is meaningful only in relation to the range of that attribute being evaluated. A change in serious adverse-event risk from 12% to 2% is not the same preference question as a change from 4% to 2%. Swing weighting makes the range explicit: it asks how valuable it is to move an outcome from its worst to its best clinically plausible level, relative to making that complete move for other outcomes.
This approach is useful when a treatment decision involves outcomes with different units, such as response probability, adverse-event risk, and symptom burden. It also makes clear why a weight cannot be interpreted independently of its anchors. If the plausible range of an outcome changes, the full swing changes, and its relative weight may need to be elicited again.
Defining the Swings and Eliciting Ratings
Begin by specifying the decision and the attributes to be considered. For each attribute, define the outcome scale, which direction is better, and a clinically plausible worst and best level. These are decision-context anchors, not necessarily the smallest and largest values observed in a particular data set. They should be chosen with clinical and patient input and should be held fixed when comparing the treatments in that decision.
Next, imagine a reference profile in which every attribute is at its worst level. Ask the decision maker to compare the individual full swings away from that profile: which improvement would matter most, and how large is each other improvement relative to it? A common elicitation format assigns 100 points to the most valuable swing and asks for ratings for the remaining swings. These are relative preference ratings, not probabilities, percentages of patients, or treatment effects. Interpreting a rating of 60 as a swing worth 60% of the reference swing assumes that the person can make a meaningful cardinal comparison on that scale; it is not justified by ranking alone.
Once the ratings are elicited, define a value function for each attribute. Linear value functions are often used when a constant value per unit change is acceptable. If the preference for intermediate levels is nonlinear, a different function may be used, but it must preserve the same endpoint anchors for the swing-weight interpretation: \(v_j(L_j)=0\) and \(v_j(U_j)=1\). For example, if a nonlinear function instead gives endpoint values 2 and 5, the full swing changes \(q_jv_j\) by \(3q_j\), not \(q_j\); the raw \(q_j\) would no longer represent the full swing as defined here.
State the clinical meaning, units, direction of preference, and plausible worst and best levels for every outcome.
Imagine each attribute moving from worst to best while all other attributes stay at worst; rank and then rate these improvements on a common preference scale.
Assign value 0 to each worst level and value 1 to each best level; specify how intermediate levels map to values.
Multiply each attribute’s value by its swing rating, add the contributions, and compare the resulting scores.
Why the Weighted Score Represents a Swing Comparison
The additive score has a direct interpretation when the model’s assumptions hold. Consider the reference profile with every outcome at its worst level. Since each value function equals zero at that level, its score is zero. Now change only attribute \(j\) from worst to best. Its value changes from 0 to 1, while all other contributions remain zero. The score therefore changes by exactly \(q_j\).
For two treatment profiles \(A\) and \(B\), subtracting their scores gives the weighted comparison:
Each term is that attribute’s contribution to the score difference. A positive total favors \(A\) according to the specified preferences and model; it does not establish that \(A\) is clinically superior for every patient or that the inputs are known without uncertainty.
This interpretation relies on important assumptions. First, the value functions correctly represent preferences over levels, including any curvature between the endpoints. Second, the additive form is suitable: there are no important interactions such that the value of an improvement in one outcome depends on the level of another. Third, the ratings compare full swings using a consistent preference scale, with the same clinical ranges and context for all treatments. If these assumptions fail, the arithmetic can still be performed, but the score may not represent the intended preference judgment.
Worked Example: Rating Three Clinical Outcome Swings
Suppose a hypothetical preference-elicitation exercise considers three attributes: response probability from 30% to 70%, serious adverse-event risk from 12% to 2%, and monthly symptom days from 12 to 4. The directions are clear: higher response probability is better, while lower adverse-event risk and fewer symptom days are better.
The decision maker judges the full response-probability improvement to be the most valuable swing and assigns it 100 points. After comparing each other full swing with that reference, the decision maker assigns 70 points to reducing serious adverse-event risk and 45 points to reducing symptom days. Thus the elicited ratings are \(q_{\text{response}}=100\), \(q_{\text{risk}}=70\), and \(q_{\text{symptoms}}=45\). These numbers encode stated relative preferences; they are not estimated from the outcome data.
For linear value functions, the normalized values are calculated from each attribute’s own endpoints. For response probability \(r\), \(v_{\text{response}}(r)=(r-0.30)/(0.70-0.30)\). For adverse-event risk \(a\), where lower is better, \(v_{\text{risk}}(a)=(0.12-a)/(0.12-0.02)\). For symptom days \(d\), \(v_{\text{symptoms}}(d)=(12-d)/(12-4)\). Each formula gives 0 at that attribute’s worst anchor and 1 at its best anchor.
For example, at a hypothetical profile with response probability 50%, adverse-event risk 6%, and 8 symptom days, the values are:
The score is \(100(0.500)+70(0.600)+45(0.500)=50+42+22.5=114.5\) swing points. The score is a composite preference score under this particular set of ranges and ratings; it is not a clinical measurement with units such as probability or days.
Worked Example: Comparing Two Hypothetical Treatments
Use the same hypothetical anchors and ratings. Suppose treatment A has response probability 60%, adverse-event risk 6%, and 8 symptom days. Treatment B has response probability 50%, adverse-event risk 3%, and 10 symptom days. The linear value functions are those defined in the preceding example.
For A, the values are \(v_{\text{response}}=(0.60-0.30)/0.40=0.750\), \(v_{\text{risk}}=(0.12-0.06)/0.10=0.600\), and \(v_{\text{symptoms}}=(12-8)/8=0.500\). Therefore \(S(A)=100(0.750)+70(0.600)+45(0.500)=75+42+22.5=139.5\) swing points.
For B, the values are \(v_{\text{response}}=(0.50-0.30)/0.40=0.500\), \(v_{\text{risk}}=(0.12-0.03)/0.10=0.900\), and \(v_{\text{symptoms}}=(12-10)/8=0.250\). Therefore \(S(B)=100(0.500)+70(0.900)+45(0.250)=50+63+11.25=124.25\) swing points.
The difference is \(S(A)-S(B)=139.5-124.25=15.25\) swing points. Its attribute contributions are \(100(0.750-0.500)=25\) for response, \(70(0.600-0.900)=-21\) for risk, and \(45(0.500-0.250)=11.25\) for symptoms. They sum to \(25-21+11.25=15.25\). Under the elicited preferences, anchors, value functions, and additive assumptions, A has the higher score. This result is conditional on those choices and does not by itself account for uncertainty in clinical outcomes or preferences.
Worked Example: Using a Nonlinear Value Function Correctly
Suppose a hypothetical decision maker values response-probability improvements more strongly near the lower end of the plausible range. Keep the response anchors at 30% and 70% and the swing rating at 100 points, but use \(v(r)=\sqrt{(r-0.30)/0.40}\). This function is nonlinear between the anchors, yet \(v(0.30)=\sqrt{0}=0\) and \(v(0.70)=\sqrt{1}=1\), so it retains the required endpoint normalization.
At response probability 50%, the value is \(v(0.50)=\sqrt{(0.50-0.30)/0.40}=\sqrt{0.5}=0.7071\). Its score contribution is \(100(0.7071)=70.71\) swing points. At the full swing endpoint of 70%, the contribution is \(100(1)=100\) points; at 30%, it is \(100(0)=0\). Thus the nonlinear shape changes the value of an intermediate outcome, but not the interpretation of the full swing rating. Using a function whose endpoint values were not 0 and 1 would require changing the scoring model or the interpretation of \(q\).
Diagnostics, Alternatives, and Limitations
A calculated score can look precise even when its preference inputs are unstable. Review the elicitation and the model rather than relying only on the final ranking.
- Check the anchors. Confirm that each endpoint is clinically plausible and describes the same decision context for each treatment. Observed sample extrema may be poor anchors, particularly when they are driven by a small or unusual sample.
- Check the value-function endpoints. For every attribute, verify \(v_j(L_j)=0\) and \(v_j(U_j)=1\). For a lower-is-better outcome, confirm that the formula assigns the correct endpoint values rather than reversing the preference direction.
- Check the elicitation. Ask whether the decision maker understood that each rating concerns a complete swing from worst to best, not a one-unit change. Revisit ratings that conflict with the stated ranking or that are difficult to distinguish reliably.
- Check the additive assumption. Discuss whether a benefit is valued differently at different levels of harm. A substantial interaction can make a simple sum misleading.
- Check sensitivity. Recalculate the comparison under plausible alternative ratings, ranges, and intermediate value-function shapes. If a small, defensible change reverses the ranking, report the decision as preference-sensitive rather than presenting one score as decisive.
Swing weighting is different from assigning weights to attributes without first specifying their ranges. The latter can obscure what improvement each weight represents. It is also different from ranking swings alone: rankings identify order but do not provide the relative magnitudes needed for a weighted sum. Direct numerical ratings are more informative for scoring, but they demand stronger judgments about how large one swing is relative to another. Pairwise comparisons can help structure those judgments, though they do not remove the need for coherent ranges and value functions.
The method has limits. Ratings may vary across patients or stakeholder groups, and a single set of values can conceal that variation. Chosen anchors can substantially influence the result, so a swing weight is not a context-free property of an outcome. Finally, a score combines preference inputs with clinical outcome values; it should not be treated as a substitute for examining the individual outcomes, the evidence supporting them, or the consequences of model assumptions.
Check Your Understanding
Answer each question, then check the solutions below.
- Why must a swing-weight elicitation specify plausible worst and best levels for every attribute?
- A value function has \(v(L)=0.2\) and \(v(U)=0.8\), with swing rating \(q=50\). How much does its contribution change from \(L\) to \(U\), and does this match the swing-weight interpretation used here?
- For an outcome where lower is better, a treatment changes the value from 0.4 to 0.7. If its swing rating is 60 points, what is its contribution to the score difference?
- What does a higher additive score establish, and name one assumption that could undermine its interpretation.
Solutions. 1. The weight refers to the value of a complete swing, so without endpoints its meaning and comparability with other swings are unclear. 2. The change is \(50(0.8-0.2)=30\) points, not 50. It does not match the defined interpretation because the function is not normalized to 0 and 1; re-anchor the function or revise the scoring model and weight interpretation. 3. The contribution is \(60(0.7-0.4)=18\) points in favor of the treatment with value 0.7. The fact that lower is better is already reflected in the value scale. 4. It establishes that the treatment has a higher score under the specified preferences, ranges, value functions, and model. For example, if the value of one outcome depends strongly on the level of another, the additive assumption may fail.