Setting Benefit-Risk Weights for a Treatment Decision

From Modeled Outcomes to a Treatment Choice

A model can predict how likely clinical benefit and serious harm are at each dose, but those probabilities alone do not determine which dose to choose. A patient who places a high value on avoiding a serious harm may prefer a lower dose; someone who places greater value on the chance of benefit may prefer a higher one. A transparent decision analysis makes this trade-off explicit by assigning weights to the outcomes and comparing the resulting expected utility across options.

The weights are not properties of the dose or the clinical model. They represent how much the decision-maker values the specified outcomes. The analysis therefore does not discover one universally correct dose. It shows how a preferred option depends on the probabilities and the values used in the comparison.

Definition: Let $p_{B,j}$ be the probability of a defined clinically meaningful benefit under option $j$, and $p_{H,j}$ the probability of a defined serious harm over the same time horizon. Let $w_B>0$ be the utility value assigned to benefit and $w_H>0$ the disutility assigned to serious harm. Under an additive outcome-utility model, the expected-utility score for option $j$ is $$ U_j=w_Bp_{B,j}-w_Hp_{H,j}. $$ The preferred option is one with the largest score among those being considered. The ratio $\lambda=w_H/w_B$ expresses the relative weight on harm; dividing every score by $w_B$ gives the equivalent normalized score $S_j=p_{B,j}-\lambda p_{H,j}$, with $\lambda>0$.

Assumptions Behind the Weighted Score

The formula is simple, but its interpretation depends on clear definitions and assumptions. The benefit and harm probabilities must refer to the same target population and a common, clinically relevant time horizon. Their event definitions should be specified in advance: for example, what qualifies as a meaningful response and which events count as serious harm. The probabilities should come from evidence or a model appropriate to the decision, not from incomparable populations or follow-up periods.

The additive score assigns the same value to a benefit and the same cost to a harm regardless of dose or whether the other event occurs. Formally, it corresponds to assigning an individual utility of $w_B I_B-w_H I_H$, where $I_B$ and $I_H$ indicate whether benefit and harm occur. Taking an expectation yields $w_BP(B)-w_HP(H)$, whether or not the two events are statistically independent. The model does not include other outcomes, interactions between benefit and harm, or differences in how outcomes are valued across patients unless these are explicitly added.

The options must also be meaningfully comparable. In the hypothetical case below, standard care and each dose are mutually exclusive strategies for the same decision. The probability estimates are treated as known for the arithmetic; in an actual application, they are estimates with uncertainty. The score is a decision aid conditional on its inputs, not a substitute for checking the evidence that produced them.

A Hypothetical Dose-Selection Case

Suppose a model provides the probabilities below for a hypothetical patient population over one year. The clinical benefit is a prespecified meaningful improvement, and serious harm is a prespecified serious adverse outcome. The numbers are illustrative only.

Option	Probability of benefit $p_B$	Probability of serious harm $p_H$
Standard care	0.30	0.02
Low dose	0.60	0.04
Middle dose	0.70	0.07
High dose	0.75	0.13

Both probabilities increase in this illustration as dose increases: higher doses bring greater modeled benefit probability but also greater serious-harm probability. That pattern does not itself determine the choice. The decision depends on the relative weight placed on benefit and harm.

Worked Example: Comparing All Options at Equal Weights

Suppose the decision-maker assigns equal magnitude to benefit and serious harm, so $w_B=w_H$ and $\lambda=1$. Calculate the normalized score for each option using $S_j=p_{B,j}-\lambda p_{H,j}$.

Standard care: $S=0.30-(1)(0.02)=0.28$.

Low dose: $S=0.60-(1)(0.04)=0.56$.

Middle dose: $S=0.70-(1)(0.07)=0.63$.

High dose: $S=0.75-(1)(0.13)=0.62$.

The middle dose has the largest score, $0.63$, so it is preferred under these probabilities and this weight ratio. This does not mean the middle dose is best for every patient or under every reasonable valuation; changing $\lambda$ can change the ranking.

Deriving the Preference Thresholds

Each option has a score that is a straight line as a function of $\lambda$: its intercept is the benefit probability and its slope is minus the harm probability. A crossover occurs where two options have equal scores. If options $i$ and $j$ have different harm probabilities, solving $p_{B,i}-\lambda p_{H,i}=p_{B,j}-\lambda p_{H,j}$ gives the exact crossover

\lambda=\frac{p_{B,i}-p_{B,j}}{p_{H,i}-p_{H,j}}.

This expression identifies a positive crossover only when the numerator and denominator have the same sign. If the harm probabilities are equal, the scores are parallel: the option with the larger benefit probability has the larger score at every $\lambda$. A crossover alone does not prove either option is best overall, because a third option could score higher. The full set of scores must be compared.

Worked Example: Finding the High-to-Middle Dose Crossover

For the high dose, $S_H=0.75-0.13\lambda$. For the middle dose, $S_M=0.70-0.07\lambda$. Set the scores equal and solve:

0.75-0.13\lambda=0.70-0.07\lambda

Subtracting $0.70$ and adding $0.13\lambda$ to both sides gives $0.05=0.06\lambda$. Thus the exact crossover is $\lambda=0.05/0.06=5/6$, approximately $0.8333$. The high dose scores higher than the middle dose for $\lambda<5/6$; the middle dose scores higher for $\lambda>5/6$; and they tie at exactly $\lambda=5/6$. The exact fraction, not its rounded display, defines the boundary.

The other adjacent crossovers are also exact. For middle versus low, the benefit-probability difference is $0.70-0.60=0.10$, and the harm-probability difference is $0.07-0.04=0.03$, giving $0.10/0.03=10/3$. For low dose versus standard care, the corresponding differences are $0.60-0.30=0.30$ and $0.04-0.02=0.02$, giving $0.30/0.02=15$.

Result: For these hypothetical probabilities and the domain $\lambda>0$, the high dose is preferred when $0<\lambda<5/6$; the middle dose when $5/6<\lambda<10/3$; the low dose when $10/3<\lambda<15$; and standard care when $\lambda>15$. The adjacent options tie at the exact crossover values $5/6$, $10/3$, and $15$, respectively. At each crossover, that tied score is at least as high as the scores of the other options.

Worked Example: Checking the Full Ranking at a High Harm Weight

Suppose $\lambda=5$, meaning that the assigned disutility of serious harm is five times the assigned utility value of benefit. The normalized scores are:

Standard care: $0.30-(5)(0.02)=0.20$.

Low dose: $0.60-(5)(0.04)=0.40$.

Middle dose: $0.70-(5)(0.07)=0.35$.

High dose: $0.75-(5)(0.13)=0.10$.

The low dose is preferred, with score $0.40$. This agrees with the exact interval $10/3<5<15$. Checking every option matters: comparing only high and low dose would not establish whether standard care or the middle dose had a higher score.

Using the Analysis and Checking Its Limits

A useful presentation reports both the probability inputs and how the preferred option changes over a plausible range of weights. For this example, the thresholds show precisely which values of $\lambda$ favor each strategy. Decision-makers can then assess whether those ranges are compatible with the preferences relevant to their decision. A range of weights is a sensitivity analysis, not an estimate of how many patients hold each preference.

The arithmetic is conditional on the modeled probabilities. Before interpreting the ranking, check that probability estimates are on the same time horizon, are defined for the same population, and correspond to the same benefit and harm definitions across options. Investigate whether sparse data, model extrapolation, or a poorly supported dose-response pattern could materially change them. Where probability estimates are uncertain, show how that uncertainty affects the scores or crossover values rather than presenting a single ranking as certain.

A weighted score is not the only decision approach. An unweighted comparison of benefit and harm probabilities is descriptive but does not resolve their trade-off. A threshold rule can instead require, for example, that serious-harm probability remain below a specified limit; that is useful when a safety constraint is non-negotiable, but it is not the same as compensating harm against benefit with a weight. More elaborate utility models can represent additional outcomes, nonlinear preferences, or patient-level variation, but require more inputs and assumptions. The simple weighted model is most useful when its two outcomes capture the central trade-off and its weights can be stated and defended.

Common pitfall: A reader may see that the high dose has the greatest benefit probability and call it the preferred dose. That conclusion ignores its greater serious-harm probability and silently assumes a particular value for harm. The correction is to state the weights, calculate the score for every option, and describe the result as conditional on those weights and modeled probabilities. Another error is to use a rounded crossover, such as $0.8333$, as if it were an exact boundary. Here the exact value is $5/6$; use the fraction for strict ranges and identify the decimal only as an approximation.

Check Your Understanding

Answer each question using the hypothetical probabilities in the table. Solutions follow the questions.

At $\lambda=0.5$, calculate all four normalized scores and identify the preferred option.
At $\lambda=4$, which option is preferred? Verify it by calculating all four scores.
What is the exact crossover between the middle and low doses? State which dose scores higher on either side.
At $\lambda=15$, which options tie, and what is their common normalized score?
Why does a higher probability of benefit alone not establish that a dose is preferred?

Solutions. (1) Standard care: $0.30-(0.5)(0.02)=0.29$; low: $0.60-(0.5)(0.04)=0.58$; middle: $0.70-(0.5)(0.07)=0.665$; high: $0.75-(0.5)(0.13)=0.685$. The high dose is preferred. (2) At $\lambda=4$, standard care scores $0.30-0.08=0.22$, low scores $0.60-0.16=0.44$, middle scores $0.70-0.28=0.42$, and high scores $0.75-0.52=0.23$. The low dose is preferred. (3) The crossover is $(0.70-0.60)/(0.07-0.04)=0.10/0.03=10/3$. The middle dose scores higher for $\lambda<10/3$, the low dose for $\lambda>10/3$, and they tie at $10/3$. (4) The low-dose score is $0.60-(15)(0.04)=0$, and standard care's score is $0.30-(15)(0.02)=0$. They tie at zero; the other options score less. (5) A dose's preferred status depends on both benefit and harm probabilities and on their assigned relative weights. A larger benefit probability does not account for the disutility associated with serious harm.