Multiplicative Benefit-Risk Utility

Why Multiply Benefit and Safety?

A dose can improve an efficacy outcome while also increasing harm. A dose-finding objective therefore needs to express how the two dimensions combine, rather than treating either one as the entire decision. Additive utility represents a weighted sum of component scores. Multiplicative utility offers a different structure: a low score on either dimension can substantially reduce the overall score, even when the other dimension is high.

This structure is useful when the decision maker wants strong performance on both benefit and safety, and regards a shortfall in one dimension as difficult to compensate for with a gain in the other. Multiplication does not automatically make an objective clinically appropriate, however. The component scales and weights must represent a defensible decision rule, and the resulting ranking should be checked against the component outcomes.

Definition: Let $d$ denote dose. Let $B(d)>0$ be a benefit score and $S(d)>0$ a safety desirability score, with larger values on each scale representing better outcomes. A multiplicative benefit-risk utility is $$ U_{\mathrm{mult}}(d)=B(d)^{w_B}S(d)^{w_S}, $$ where $w_B>0$ and $w_S>0$ are weights. If the scores are normalized to lie between 0 and 1, then larger $U_{\mathrm{mult}}(d)$ indicates a more preferred benefit-safety profile under this specified utility.

The safety component is a desirability score, not a harm score. For example, if $R(d)$ is an adverse-event probability, one possible safety score is $S(d)=1-R(d)$, the probability of avoiding that event. The mapping from risk to safety score is part of the objective and should be stated explicitly. Multiplying benefit by an adverse-event probability would instead reward higher harm risk, which reverses the intended direction.

The formula requires positive component values when logarithms or derivatives are used. If a component can equal zero, the product is zero regardless of the other component. That may be a deliberate consequence of the decision rule, but it can also make a utility unusable for optimization or comparisons. A zero should not be replaced by a small positive number without explaining how that changes the objective.

Interpreting the Multiplicative Structure

Taking logarithms gives

\log U_{\mathrm{mult}}(d) =w_B\log B(d)+w_S\log S(d).

Thus the multiplicative rule is additive on the log-score scale, not on the original benefit and safety scales. A change in a component matters in proportion to its relative change: differentiating gives $d\log B=dB/B$. Consequently, the same absolute increase in benefit has a larger effect on log utility when the starting benefit score is small than when it is already large, all else equal.

The ratio of the exponents determines the relative weighting of the two log components. Multiplying both exponents by the same positive constant raises the utility to a positive power and therefore does not change dose rankings. If $w_B+w_S=1$, the formula is a weighted geometric mean; this normalization keeps the utility between zero and one when both component scores are in that range. The sum-to-one convention is convenient, but it is not necessary for ranking doses.

Component scales matter. A score of 0.8 must have a defensible meaning relative to another value on the same component scale; a probability and a response measured in milligrams cannot simply be multiplied as raw quantities and treated as a meaningful utility. Scores should be defined as dimensionless desirability values, or otherwise transformed so that their scales and directions are explicit.

Worked Example: Building a Product from Benefit and Safety Scores

Consider two hypothetical doses with normalized predicted benefit and safety scores. Dose A has $B=0.80$ and $S=0.50$. Dose B has $B=0.70$ and $S=0.70$. First use equal exponents of one, so $U_{\mathrm{mult}}=BS$.

For dose A, $$ U_{\mathrm{mult}}(A)=0.80(0.50)=0.400. $$ For dose B, $$ U_{\mathrm{mult}}(B)=0.70(0.70)=0.490. $$ The product ranks dose B higher, despite its lower benefit score, because its higher safety score more than compensates under this particular product rule. This is a ranking implied by the specified scores and formula, not a general claim that dose B is clinically preferable.

Now consider a rule that places more exponent weight on safety: $w_B=0.4$ and $w_S=0.6$. The utilities are $$ U_{\mathrm{mult}}(A)=0.80^{0.4}0.50^{0.6}\approx 0.6034,\qquad U_{\mathrm{mult}}(B)=0.70^{0.4}0.70^{0.6}=0.7000. $$ Dose B remains higher. The exponents alter the way relative component levels contribute; they do not simply multiply the benefit and safety scores by linear coefficients.

Optimizing Utility Across Dose

When $B(d)$ and $S(d)$ are differentiable and positive, maximizing their product is equivalent to maximizing its logarithm, because the logarithm is strictly increasing. An interior optimum must satisfy

\frac{d}{dd}\log U_{\mathrm{mult}}(d) =w_B\frac{B'(d)}{B(d)} +w_S\frac{S'(d)}{S(d)}=0.

This condition balances weighted relative rates of change. A positive relative change in benefit can offset a negative relative change in safety. Unlike the additive marginal tradeoff condition, the rates are divided by the current component values. This means that the tradeoff generally changes with dose even if the absolute slopes $B'(d)$ and $S'(d)$ do not.

The equation identifies a stationary point, not necessarily a maximum. To establish a maximum, also check curvature or compare utility over the admissible dose range, including its boundaries. The following example verifies the assumptions and the global maximum directly.

Worked Example: Finding the Utility-Maximizing Dose

Suppose, hypothetically, that dose $x$ is measured in milligrams and the positive, dimensionless scores are $$ B(x)=\frac{x}{1+x},\qquad S(x)=e^{-0.4x},\qquad x>0. $$ Use the equal-exponent product $U(x)=B(x)S(x)$. Both scores are positive and differentiable for $x>0$, and higher $B$ is better while $S$ declines with dose.

It is simpler to differentiate the log utility: $$ \log U(x)=\log x-\log(1+x)-0.4x, $$ so $$ \frac{d}{dx}\log U(x)=\frac{1}{x(1+x)}-0.4. $$ Setting this derivative to zero gives $x(1+x)=2.5$, or $x^2+x-2.5=0$. The positive solution is $$ x^*=\frac{-1+\sqrt{11}}{2}\approx 1.1583\ \text{mg}. $$ The second derivative is $$ \frac{d^2}{dx^2}\log U(x) =-\frac{1+2x}{x^2(1+x)^2}<0\quad\text{for }x>0. $$ Thus log utility is strictly concave on the domain. Also, $U(x)$ approaches zero as $x$ approaches zero and as $x$ grows without bound, so this stationary point is the global maximum for $x>0$.

At the maximizing dose, $B(x^*)\approx 0.5367$ and $S(x^*)\approx 0.6292$. Evaluating the product at the unrounded maximizing value gives $$ U(x^*)\approx 0.3377. $$ The result is a mathematical optimum for these hypothetical score functions, not a dose recommendation. A practical choice would also have to respect the clinically relevant dose range and the validity of the predictions.

When Rankings Differ from Additive Utility

An additive objective might use $U_{\mathrm{add}}(d)=B(d)+S(d)$ when equal linear weights are intended. The product and sum can rank the same candidates differently because the sum rewards absolute gains, whereas the product reflects relative component levels and penalizes imbalance. A rank reversal is not an inconsistency: it shows that the two formulas encode different preferences.

Worked Example: A Rank Reversal Between Sum and Product

Suppose two hypothetical dose profiles have normalized scores. Dose A has $B=0.90$, $S=0.40$; dose C has $B=0.62$, $S=0.62$. Compare the equal-weight additive sum with the equal-exponent product.

For the additive sum, $$ U_{\mathrm{add}}(A)=0.90+0.40=1.30,\qquad U_{\mathrm{add}}(C)=0.62+0.62=1.24. $$ The sum ranks A higher. For the multiplicative objective, $$ U_{\mathrm{mult}}(A)=0.90(0.40)=0.360,\qquad U_{\mathrm{mult}}(C)=0.62(0.62)=0.3844. $$ The product ranks C higher. Dose A's benefit advantage raises its sum, but its lower safety score depresses its product. Dose C's more balanced profile wins under multiplication. Neither ranking is universally correct; the appropriate rule depends on the intended consequences and preferences represented by the objective.

Assumptions, Diagnostics, and Limitations

A multiplicative ranking is interpretable only when several conditions are met. The benefit and safety scores must describe the same dose decision and target population; larger values must consistently mean better outcomes; and the weights and score mappings must be specified before interpreting the resulting ranking. For calculus-based optimization, the component functions must be positive and differentiable in the dose range, and the stationary point must be checked against the full feasible range.

Useful diagnostics include plotting $B(d)$, $S(d)$, and $U_{\mathrm{mult}}(d)$ over the dose range, and checking whether the selected dose is driven by an unexpected score boundary or an extreme component value. Repeat the ranking under plausible changes to the exponents and score mappings. If modest changes reverse the recommendation, the decision is sensitive to utility specification and should be reported as such rather than hidden behind a single optimum.

Reversed direction: If a harm probability is used directly as $S(d)$, larger harm risk increases utility. Define a safety desirability score with the correct direction.
Unexamined zeros: If either component is zero, the product is zero. Confirm that this consequence is intended and that the optimization is well-defined.
Arbitrary scales: Changing a component's numerical scale can change product rankings. Specify how scores are constructed and what their values mean.
Interpreting weights as fixed absolute tradeoffs: The exponents act on log scores. The implied absolute tradeoff varies with the current component levels.
Reporting only the winning dose: A utility maximum conceals which dimension drives the result. Report component scores alongside utility so the benefit-risk profile remains visible.

The product is one decision model, not an empirical law. It may be unsuitable when a component has a mandatory minimum, when a zero score should not erase all value in the other component, or when stakeholders prefer a different form of compensation. Such structures require a different utility specification. In particular, a multiplicative form should not be selected merely because both components are available numerically.

Key takeaway: Multiplicative utility combines positive benefit and safety desirability scores through weighted powers. Its log form makes the relative contributions explicit, but its rankings depend on the score scales and exponents. Always inspect the component predictions as well as the combined utility.

Check Your Understanding

Answer each question, then compare with the solutions below.

A safety score is defined as $S(d)=1-R(d)$, where $R(d)$ is the adverse-event probability. If $R(d)=0.25$, what is $S(d)$, and why is it preferable to multiplying benefit by $R(d)$ when higher utility should mean better safety?
For $B=0.80$, $S=0.50$, $w_B=1$, and $w_S=1$, calculate multiplicative utility.
Two candidates have scores $(B,S)=(0.80,0.50)$ and $(0.70,0.70)$. Which has greater equal-exponent product utility?
State the interior stationary-point condition for $B(d)^{w_B}S(d)^{w_S}$ when both component functions are positive and differentiable.

Solutions. 1. $S=1-0.25=0.75$; it increases as adverse-event risk falls, while using $R$ directly would reward higher risk. 2. $U=0.80(0.50)=0.400$. 3. The products are $0.80(0.50)=0.400$ and $0.70(0.70)=0.490$, so the second candidate is higher. 4. The condition is $w_B B'(d)/B(d)+w_S S'(d)/S(d)=0$; it identifies a stationary point, which must still be checked as a maximum over the admissible range.