Additive Multi-Criteria Benefit-Risk Models

Why Combine Multiple Clinical Attributes?

A treatment decision may involve several outcomes at once. A therapy could improve the chance of clinical response while also increasing the chance of toxicity. A single efficacy endpoint or a single harm measure cannot, by itself, represent the whole trade-off. An additive multi-criteria model makes that trade-off explicit by converting each attribute to a value on a defined scale and combining the values with preference weights.

The resulting score is a structured summary of a specified treatment profile under specified preferences. It is not a probability of success, a treatment effect in clinical units, or proof that one treatment is clinically superior for every patient. Its meaning depends on the attributes, value functions, weights, and assumptions used to construct it.

Definition: For a profile with $J$ clinical attributes $x_1,\ldots,x_J$, an additive multi-criteria value score is $$ V(x_1,\ldots,x_J)=\sum_{j=1}^{J} w_j v_j(x_j), $$ where $v_j(x_j)$ is the value assigned to attribute $j$, and $w_j$ is its preference weight. A common convention is to scale each value function from 0 (least preferred anchor) to 1 (most preferred anchor), with $w_j\geq 0$ and $\sum_{j=1}^{J}w_j=1$. Under this convention, $V$ lies between 0 and 1.

The attribute-specific value function determines how a clinical measurement maps to preference. Its direction must make sense: if a higher response probability is preferable, its value should not decrease as that probability increases; if a higher toxicity risk is worse, its value should not increase as the risk increases. The function need not be linear. A weight describes the relative importance of a full swing across the chosen value-function range; it is not the importance of one unit of the raw clinical measurement.

Construction and Assumptions

Start by specifying the decision and the attributes to include. Define the clinically relevant range and anchors for each attribute, then construct a value function that reflects how changes across that range matter to the decision. Finally, elicit or otherwise justify the weights for those full-range changes. Weights that sum to one are convenient for interpretation, but the normalization alone does not validate the preferences or the model.

The additive form has a substantive assumption: the value of a change in one attribute can be represented separately from the levels of the other attributes. In preference theory, a form of preferential independence is needed for an additive representation. In practical terms, the trade-off between efficacy and toxicity should not change in an unmodeled way depending on the level of another included attribute. If, for example, the value of avoiding a toxicity depends strongly on whether efficacy is high or low, a simple additive score may not represent the decision-maker’s preferences adequately.

Other assumptions also matter. The attributes should cover the outcomes relevant to the decision without double-counting the same consequence. The anchors and value functions should be appropriate to the population and decision context. Weights should be elicited for those specific value-function ranges, rather than copied from a different set of ranges. For treatment outcomes that are uncertain, the score must be based on a clearly defined profile or on values averaged over the outcome distributions; a score for one profile should not be presented as though it automatically represents all possible outcomes.

Worked Clinical Comparison

The following examples use hypothetical treatment profiles. They illustrate model arithmetic, not evidence about actual treatments. In the first comparison, the stated probabilities are treated as the profiles to be scored. The anchors and weights are assumed to have been judged appropriate for the decision context.

Worked Example: Scoring Efficacy and Toxicity Together

Suppose the decision-maker values response probability from 30% (value 0) to 70% (value 1), and values severe-toxicity probability from 30% (value 0) to 5% (value 1). Between the anchors, use linear value functions. Assign efficacy weight 0.65 and toxicity weight 0.35. Hypothetical treatment A has a 58% response probability and 10% severe-toxicity probability; treatment B has corresponding probabilities of 62% and 15%.

For response probability $p$, the value function is $v_E(p)=(p-0.30)/(0.70-0.30)$. For severe-toxicity probability $q$, lower is better, so $v_T(q)=(0.30-q)/(0.30-0.05)$. These functions map their stated least- and most-preferred anchors to 0 and 1.

For treatment A, the attribute values are $$ v_E(0.58)=\frac{0.58-0.30}{0.40}=0.700,\qquad v_T(0.10)=\frac{0.30-0.10}{0.25}=0.800. $$

Its additive score is $$ V_A=0.65(0.700)+0.35(0.800)=0.455+0.280=0.735. $$

For treatment B, $$ v_E(0.62)=\frac{0.62-0.30}{0.40}=0.800,\qquad v_T(0.15)=\frac{0.30-0.15}{0.25}=0.600, $$ so $$ V_B=0.65(0.800)+0.35(0.600)=0.520+0.210=0.730. $$

Under these value functions and weights, A scores 0.005 value units higher than B. This small difference does not establish a meaningful clinical advantage: its practical importance depends on the decision context and the credibility of the inputs. The scores express preference-weighted values on the specified scale, not response probabilities.

How Attribute Changes Contribute to a Score

The additive form makes the contribution of each attribute explicit. For two profiles, A and B, subtracting their scores gives a sum of weighted attribute-value differences. This is useful for checking calculations and understanding which trade-offs determine the comparison.

V_B-V_A=\sum_{j=1}^{J}w_j\left[v_j(x_{Bj})-v_j(x_{Aj})\right].

This follows by subtracting the two additive definitions and collecting the terms for each attribute:

\begin{aligned} V_B-V_A &=\sum_{j=1}^{J}w_jv_j(x_{Bj})-\sum_{j=1}^{J}w_jv_j(x_{Aj})\\ &=\sum_{j=1}^{J}w_j\left[v_j(x_{Bj})-v_j(x_{Aj})\right]. \end{aligned}

A positive difference means B has the higher score under the specified model; a negative difference means A has the higher score. Neither sign is a probability statement. Each difference is conditional on the selected attributes, anchors, value functions, and weights.

Worked Example: Checking a Weighted Trade-Off

Suppose two attributes have the same value-function swing of 0.15 in their original value units. Moving from A to B improves the efficacy value by 0.08 units but worsens the toxicity value by 0.11 units. After dividing each change by its 0.15 swing, the unweighted sum of the standardized changes is $$ \frac{0.08}{0.15}-\frac{0.11}{0.15} =0.5333-0.7333=-0.2000. $$

With equal weights of 0.50, the model’s weighted score change is $$ V_B-V_A=0.50\left(\frac{0.08}{0.15}\right) +0.50\left(-\frac{0.11}{0.15}\right) =0.2667-0.3667=-0.1000. $$

Thus B’s score is 0.1000 value units below A’s under equal weights. The unweighted standardized sum is $-0.2000$; applying the equal weights gives the model score difference of $-0.1000$. Keeping these two calculations distinct avoids confusing a raw sum of standardized changes with a weighted score.

Weight Sensitivity and a Break-Even Point

A treatment ranking can depend on the weights. Even if the attribute values are fixed, a different relative emphasis on efficacy and toxicity may change which score is higher. A break-even calculation shows where that change occurs.

Worked Example: Finding the Efficacy Weight at Which Scores Tie

Consider hypothetical profiles with efficacy values of 0.75 for A and 0.90 for B, and toxicity values of 0.85 for A and 0.55 for B. Let the efficacy weight be $w$, so the toxicity weight is $1-w$. The score difference is $$ V_B-V_A=w(0.90-0.75)+(1-w)(0.55-0.85) =0.15w-0.30(1-w). $$

At a tie, $V_B-V_A=0$. Solving gives $$ 0.15w-0.30+0.30w=0,\qquad 0.45w=0.30,\qquad w=\frac{0.30}{0.45}=\frac{2}{3}\approx 0.6667. $$

If the efficacy weight exceeds $2/3$, B has the higher score; if it is below $2/3$, A has the higher score; at $2/3$ the scores tie (approximately 0.6667). For example, with $w=0.70$, the difference is $0.15(0.70)-0.30(0.30)=0.105-0.090=0.015$, so B’s score is higher by 0.015 value units. This threshold describes sensitivity to the chosen preference weight; it does not identify which weight is correct.

Diagnostics, Alternatives, and Limitations

Before using a score to support a decision, check both its arithmetic and its structure. Confirm that each value function maps the intended worst and best anchors to the stated values, that harmful attributes run in the correct direction, and that weights are nonnegative and sum to one when that convention is used. Recalculate the treatment difference attribute by attribute. A surprising result may reveal a reversed toxicity scale, inconsistent anchors, an omitted attribute, or an arithmetic error.

Then examine the assumptions. Ask whether the attributes overlap, whether their value functions reflect plausible preferences across their full ranges, and whether the trade-off between two attributes would change at different levels of a third. If rankings change under small, reasonable changes to weights or value-function shapes, report that sensitivity rather than presenting one ranking as definitive. The score also inherits uncertainty in the clinical outcome estimates and in preference judgments; a single numerical score does not make those uncertainties disappear.

An additive model differs from several alternatives. A single endpoint gives a direct comparison on one outcome but may omit important benefits or harms. A composite endpoint combines outcomes according to its event definition and analysis, which need not represent the decision-maker’s preference trade-offs. A weighted net clinical benefit can be interpreted as an additive score when its components are put on compatible value scales and its weights have a defensible meaning. A multiplicative model is a different way to represent interactions among attributes; it may be considered when additive independence is not credible, but it requires its own assumptions and construction.

Approach	What it summarizes	Key caution
Additive multi-criteria score	Weighted values across specified attributes	Depends on value functions, weights, and an additive preference structure
Single endpoint	One selected clinical outcome	Other relevant benefits and harms may be left out
Composite endpoint	A prespecified combination of events	Event composition and analysis weights may not match patient preferences
Multiplicative utility	A non-additive representation that can allow attribute interactions	Requires a different model and additional preference assumptions

Key takeaway: An additive score is meaningful only together with its attribute definitions, value-function anchors, weights, and independence assumptions. Report those ingredients alongside the score, and describe a close or weight-sensitive ranking as such.

Common Mistakes

Adding raw clinical measurements. Response probability and toxicity risk have different meanings and directions. Define attribute-specific value functions before combining them.
Treating weights as universal constants. A weight applies to the particular swing defined by its value function and anchors; changing the range can change what the weight means.
Interpreting the score as a probability. A score of 0.73 is not a 73% chance of benefit. It is a preference-weighted value on the stated scale.
Assuming a higher score proves a treatment effect. The score compares specified profiles under a model; the clinical evidence and uncertainty behind those profiles must be assessed separately.
Ignoring interactions or double-counting. If attributes overlap or their trade-offs depend strongly on other attributes, an additive total can misrepresent preferences.

Check Your Understanding

Try each problem, then compare your reasoning with the solutions below.

A toxicity value function assigns 0 to a 20% risk and 1 to a 5% risk, with linear interpolation. What value does it assign to a 11% risk?
A two-attribute model uses weights 0.60 for efficacy and 0.40 for toxicity. Treatment A has values 0.70 and 0.80; treatment B has values 0.85 and 0.55. Calculate both scores and state which scores higher.
Why is a score of 0.60 not automatically a 60% probability that treatment will work?
If the trade-off between efficacy and toxicity changes substantially depending on a third attribute’s level, what additive-model assumption is in question?

Solutions. (1) The value is $(0.20-0.11)/(0.20-0.05)=0.09/0.15=0.600$. (2) $V_A=0.60(0.70)+0.40(0.80)=0.42+0.32=0.740$. $V_B=0.60(0.85)+0.40(0.55)=0.51+0.22=0.730$. A scores higher by 0.010 value units. (3) The score is a weighted value on the chosen attribute scales, not a probability model; its numerical scale does not imply a chance of response. (4) The preferential-independence condition supporting additive representation is in question, because the efficacy-toxicity trade-off is not stable across levels of another attribute.