Why Model Variation in Preferences?
A benefit-risk comparison can depend on how patients value different outcomes. One patient may place greater weight on improved symptom control; another may be more concerned about avoiding a serious adverse event. A single set of preference weights can summarize one perspective, but it cannot show how often patients with different preferences favor different treatment profiles.
This tutorial represents between-patient variation in preference weights. The treatment outcome scores are held fixed for each comparison: the question is how the same outcome trade-off is valued by different patients. This is distinct from variation in outcomes between patients. It is also distinct from uncertainty about the estimated weights. The introductory material on patient preference weights and normalized weights provides the context for interpreting weights on a common scale.
From Preference Weights to a Treatment Preference
Consider two attributes: a benefit score and a harm-avoidance score. Suppose both attribute scores are oriented so that higher is better, and both use scales appropriate for the additive utility model. Let \(w\) be a patient's weight on benefit, so the weight on harm avoidance is \(1-w\). Thus \(0\leq w\leq 1\). For treatment \(A\) compared with treatment \(B\), let \(\delta_B\) be the difference in benefit scores and \(\delta_H\) the difference in harm-avoidance scores, each calculated as the score for \(A\) minus the score for \(B\). The patient's utility contrast is
Under this specified additive model, the patient favors \(A\) when \(\Delta U(w)>0\), favors \(B\) when \(\Delta U(w)<0\), and is indifferent when the contrast is zero. These are comparisons of utility scores for the stated treatment profiles; they do not by themselves establish that either treatment causes better outcomes.
When \(\delta_B>\delta_H\), the contrast increases with \(w\). Solving \(\Delta U(w)>0\) gives the preference threshold
If \(w^*\) lies between 0 and 1, patients above the threshold favor \(A\), and those below it favor \(B\). If \(w^*\leq 0\), all allowed weights favor \(A\), apart from any boundary tie; if \(w^*\geq 1\), all allowed weights favor \(B\), apart from a possible tie at \(w=1\). When \(\delta_B=\delta_H\), the contrast is constant in \(w\), so changing the preference weight cannot change the ranking.
Worked Example: Preference Share Under a Continuous Distribution
Suppose two hypothetical treatments have differences \(\delta_B=0.12\) and \(\delta_H=-0.08\) utility points, with both differences defined as treatment \(A\) minus treatment \(B\). Thus \(A\) scores better on benefit but worse on harm avoidance. Suppose patient benefit weights follow a Beta(2, 2) distribution, which is symmetric on the interval from 0 to 1.
The individual contrast is
Treatment \(A\) is favored when \(0.20w-0.08>0\), or \(w>0.40\). For a Beta(2, 2) random weight, the cumulative distribution function is \(F(w)=3w^2-2w^3\). Therefore,
Under the assumed preference distribution, 64.8% of patients favor \(A\) for these fixed profiles, and 35.2% favor \(B\). Because the distribution is continuous, the probability of exactly \(w=0.40\), and hence exact indifference, is zero. The 64.8% is a model-based population share, not a guarantee that exactly 64.8 of every 100 sampled patients will favor \(A\).
Continuous Distributions and Latent Classes
For two normalized weights, a distribution on \(w\) between 0 and 1 is a direct way to represent heterogeneity. A Beta distribution is one option: its parameters control the distribution's shape, allowing preferences to concentrate near the middle, near an endpoint, or across a wider range. For more than two attributes, a common representation is a Dirichlet distribution for a vector of nonnegative weights that sum to 1. The chosen distribution should be motivated by preference data and the target population, rather than selected only because it is mathematically convenient.
A latent-class model instead represents preferences as a finite mixture. Let class \(k\) have population proportion \(\pi_k\), with class-specific weights \(w_k\), where \(\pi_k\geq 0\) and \(\sum_k\pi_k=1\). Under deterministic utility comparisons, the preference share for \(A\) is the sum of the proportions of classes whose utility contrast is positive:
where \(I\{\cdot\}\) equals 1 when its condition is true and 0 otherwise. A class mixture is useful when distinct preference profiles are meaningful or supported by the data. It does not imply that each class can be identified perfectly for an individual patient; class membership may be unobserved or uncertain.
Worked Example: Preference Shares From Three Latent Classes
Consider the same hypothetical treatment differences, \(\delta_B=0.12\) and \(\delta_H=-0.08\), so that \(\Delta U(w)=0.20w-0.08\). Suppose the population has three preference classes: 35% with \(w=0.80\), 45% with \(w=0.50\), and 20% with \(w=0.20\).
For the first class, \(\Delta U(0.80)=0.20(0.80)-0.08=0.080\), so this class favors \(A\). For the second, \(\Delta U(0.50)=0.100-0.080=0.020\), so it also favors \(A\). For the third, \(\Delta U(0.20)=0.040-0.080=-0.040\), so it favors \(B\). The resulting shares are
The population mean weight is \(0.35(0.80)+0.45(0.50)+0.20(0.20)=0.545\). Its utility contrast is \(0.20(0.545)-0.08=0.029\) utility points. That positive contrast describes the utility at the mean weight; it does not mean that every patient favors \(A\). In this example, one class, comprising 20% of the population, favors \(B\).
Comparing Patient Groups
Preference shares can also be calculated separately for patient groups, such as groups defined by a prespecified clinical or demographic characteristic. Each group may have a different preference distribution or different class proportions. The comparison then asks whether the modeled treatment ranking varies across groups because their weights differ. Group labels alone do not establish why preferences differ, and estimated group differences should not be interpreted as causal effects of group membership.
Worked Example: Contrasting Preference Shares Across Two Groups
Continue to use the threshold \(w>0.40\) for favoring \(A\). Suppose, hypothetically, that group 1 has weights distributed as Beta(2, 5), while group 2 has weights distributed as Beta(5, 2). These distributions represent different patterns of preference weights on the same 0-to-1 scale.
For a Beta(2, 5) weight, the cumulative distribution function is \(F(w)=1-(1-w)^5(1+5w)\). The group 1 share favoring \(A\) is
For Beta(5, 2), symmetry around 0.5 gives \(P(w>0.40)=F_{\mathrm{Beta}(2,5)}(0.60)\). Therefore,
The modeled shares are 23.33% in group 1 and 95.90% in group 2. These results follow from the assumed distributions and treatment differences. In an application, the distributions would need to be estimated or elicited for the relevant groups; the hypothetical calculations alone do not show that any real groups differ.
Assumptions, Diagnostics, and Limitations
A preference-share calculation is interpretable only relative to its assumptions. The attribute scores must be defined on scales for which the stated utility model is meaningful, and the weights must refer to those attributes and scales. In the two-attribute example, weights are nonnegative and sum to 1. The treatment score differences must be oriented consistently, and the same preference distribution should not be applied to a different target population without justification. The calculation also assumes a deterministic ranking given a patient's weights and the stated profiles; it does not include randomness in how a patient might choose or respond.
Several checks help identify a poorly specified analysis:
- Check the direction and scale. Confirm that higher scores have the intended meaning and that each treatment difference is calculated as \(A\) minus \(B\). A reversed sign can reverse the preference conclusion.
- Check the support and normalization. Simulated or fitted two-attribute weights should lie from 0 to 1, and the two weights should sum to 1. For multiple attributes, all weights should be nonnegative and sum to 1.
- Check the threshold against the distribution. If few patients are modeled near the threshold, small changes in treatment differences may have little effect on the share. If substantial mass lies near it, small changes may shift many patients between rankings.
- Check whether the distribution is credible. A single Beta distribution may smooth over distinct groups; a class model may impose artificial boundaries or omit variation within classes. Compare model implications with elicited preferences and examine whether the fitted representation captures the observed patterns.
- Keep preference and estimation uncertainty distinct. A wide spread of patient weights describes heterogeneity. Limited data about those weights creates uncertainty about the distribution and the resulting share. A reported share should not imply that the preference distribution is known exactly.
A population mean weight is a useful summary, but it cannot generally replace a preference distribution when the question is what proportion favors each treatment. A latent-class model makes distinct profiles explicit, while a continuous distribution represents gradual variation. Neither approach is universally preferable: the choice depends on the elicitation data, model fit, and intended interpretation. Both also depend on the utility specification and the particular treatment profiles being compared. The resulting shares are not universal rankings of treatments, and preference variation alone does not incorporate uncertainty in clinical outcomes.
Check Your Understanding
Use the stated additive utility model and show how the preference share follows from the weight distribution or class proportions.
- With \(\delta_B=0.10\) and \(\delta_H=-0.05\), derive the threshold weight above which treatment \(A\) is favored.
- Using that threshold, if \(w\sim\text{Beta}(2,2)\), what proportion favors \(A\)?
- In a two-class population, 60% have \(w=0.70\) and 40% have \(w=0.20\). Using the treatment differences in question 1, calculate the shares favoring \(A\) and \(B\).
- Why does a positive utility contrast evaluated at the mean weight not establish that all patients favor \(A\)?
Solutions. For question 1, \(\Delta U(w)=0.10w-0.05(1-w)=0.15w-0.05\). Setting this greater than zero gives \(w>0.05/0.15=1/3\). For question 2, a Beta(2, 2) distribution has \(F(w)=3w^2-2w^3\), so \(P(w>1/3)=1-F(1/3)=1-(1/3-2/27)=20/27\), or 0.7407. For question 3, the contrasts are \(0.15(0.70)-0.05=0.055\) and \(0.15(0.20)-0.05=-0.020\). Thus 60% favor \(A\) and 40% favor \(B\). For question 4, the mean weight gives one utility contrast, while individual weights can fall on both sides of the threshold; the preference share requires the distribution or class proportions.