Additive Utility and the Marginal Tradeoff Condition

Why Marginal Tradeoffs Matter

When efficacy increases with dose but harm also rises, selecting the dose with the greatest efficacy alone can overlook important costs. An additive utility makes the tradeoff explicit: it assigns a contribution to benefit and a separate contribution to harm, then combines them. The central question is not whether harm exists, but whether the extra benefit from a small dose increase is worth its extra harm according to the chosen utility scale.

The marginal tradeoff condition answers that question for a smooth, continuous dose range. At an interior optimum, a small increase in dose has no first-order net utility gain: weighted marginal benefit equals weighted marginal harm. This condition identifies a candidate optimum, not automatically the best dose. The dose range, curvature, boundaries, and the validity of the modeled benefit and harm curves still matter.

Define the Additive Utility and Its Assumptions

Definition: Let $d$ be dose in a prespecified feasible interval $[d_{\min},d_{\max}]$. Let $B(d)$ denote the benefit summary and $H(d)$ the harm summary at dose $d$, defined for the same target population and decision context. An additive utility is $$ U(d)=w_B B(d)-w_H H(d), $$ where $w_B>0$ and $w_H>0$ are weights that put the benefit and harm summaries on a common utility scale. The optimal dose is any feasible dose that maximizes $U(d)$.

The summaries $B(d)$ and $H(d)$ may be predicted means, probabilities, or other prespecified summaries of outcomes. They must correspond to the same population and decision context; otherwise their difference does not represent the intended tradeoff. The weights encode how changes in each summary count toward utility, including any unit conversion. Their values must be specified for the decision, not chosen after seeing which dose they favor.

For the derivative condition, assume that the feasible interval is continuous, that $B$ and $H$ are differentiable near the dose of interest, and that the weights are fixed as dose varies. The derivatives describe local changes in the summaries per unit dose. If $B$ is measured in benefit points and dose in mg/day, for example, $B'(d)$ is measured in benefit points per mg/day. The weighted derivatives $w_BB'(d)$ and $w_HH'(d)$ must be expressed in the same utility units per dose unit.

Deriving the Marginal Tradeoff Condition

An interior optimum $d^*$ lies strictly between the feasible boundaries. If $U$ is differentiable there, a necessary condition for a local maximum is $U'(d^*)=0$. Differentiating the additive utility gives

$$ U'(d)=w_B B'(d)-w_H H'(d). $$

Therefore an interior stationary dose satisfies

$$ w_B B'(d^*)=w_H H'(d^*). $$

This is the marginal tradeoff condition. It equates weighted changes per unit dose, not the total benefit and harm at the dose. If harm increases with dose, then $H'(d)>0$; the condition says the weighted additional benefit from a small dose increase exactly balances the weighted additional harm at an interior stationary point.

A second derivative helps classify the stationary point:

$$ U''(d)=w_B B''(d)-w_H H''(d). $$

If $U''(d^*)<0$, the stationary point is a strict local maximum under the usual second-derivative condition. If $U''(d^*)>0$, it is a local minimum, not an optimum. If $U''(d^*)=0$, this test does not decide the classification. A particularly useful sufficient condition for a global solution is strict concavity of $U$ over the entire feasible interval: then any interior stationary point is the unique global maximum. Without such a condition, compare utility at all relevant stationary points and both boundaries. The derivative equation alone is not a global-optimality proof.

Worked Example: Finding a Concave Interior Optimum

Suppose a hypothetical model predicts, for doses $d$ in mg/day from 0 to 40, $B(d)=100(1-e^{-0.08d})$ benefit points and $H(d)=0.04d^2$ harm points. Let both weights equal 1 utility unit per point. Find the optimal dose.

The additive utility and its derivative are

U(d)=100(1-e^{-0.08d})-0.04d^2, \qquad U'(d)=8e^{-0.08d}-0.08d.

An interior stationary dose solves $8e^{-0.08d}=0.08d$. Numerically, the two sides are approximately 1.602 and 1.608 at $d=20.1$, so the root is about $20.1$ mg/day. The second derivative is $U''(d)=-0.64e^{-0.08d}-0.08<0$ throughout the interval. Thus $U$ is strictly concave, and this stationary dose is the unique global maximum.

At $d=20.1$, $e^{-1.608}\approx0.2003$, so $B(20.1)\approx100(1-0.2003)=79.97$ benefit points and $H(20.1)=0.04(20.1)^2=16.16$ harm points. Hence $U(20.1)\approx79.97-16.16=63.81$ utility units. The boundary utilities are $U(0)=0$ and $U(40)\approx100(1-e^{-3.2})-64=31.92$. These checks agree that the interior dose is best over the stated range.

Weights Change the Tradeoff

The same dose-response curves can yield different optimal doses when the relative weights change. The condition does not say that unweighted benefit and harm slopes are equal; it says that their weighted slopes are equal. A weight therefore has a decision interpretation: it determines how much a unit change in one summary contributes to utility relative to a unit change in the other. Before applying the condition, check that the weights and summary units actually produce comparable quantities.

Worked Example: Applying a Harm Weight

Consider a hypothetical dose $d$ measured in units of 10 mg/day, with $0\le d\le30$. Suppose $B(d)=12\ln(1+d)$ benefit points and $H(d)=0.3d$ harm points. Use $w_B=1$ utility unit per benefit point and $w_H=2$ utility units per harm point.

Then $U(d)=12\ln(1+d)-0.6d$. Its derivative is $U'(d)=12/(1+d)-0.6$. Setting the weighted marginal benefit equal to the weighted marginal harm gives

\frac{12}{1+d}=2(0.3)=0.6, \qquad 1+d=20, \qquad d=19.

The second derivative is $U''(d)=-12/(1+d)^2<0$, so this is the unique maximum over the interval. The dose is 19 units, or 190 mg/day. At this dose, $B(19)=12\ln(20)\approx35.95$ benefit points and $H(19)=0.3(19)=5.70$ harm points. The utility is $35.95-2(5.70)=24.55$ utility units. The weight of 2 makes each harm point count twice as much as a benefit point on the selected utility scale; changing that preference would change the balance condition and could change the selected dose.

Boundaries and Discrete Dose Choices

A dose must be feasible as well as attractive. If the derivative does not reach zero inside the allowed interval, the optimum may be at a boundary. For instance, with $B(d)=2d$, $H(d)=0.2d^2$, unit weights, and $0\le d\le4$, the derivative is $U'(d)=2-0.4d$, which is positive throughout the interval. Utility therefore increases across the allowed range, so the optimum is $d=4$, not the unconstrained stationary dose $d=5$. That value lies outside the feasible interval.

Some decisions allow only a dose grid rather than every real-valued dose. In that case, a derivative equation is not the selection rule: calculate utility at every allowable dose and choose the greatest. Marginal comparisons between neighboring grid points can help explain the choice, but they do not require equality of derivatives.

Worked Example: Choosing Among Three Dose Levels

Suppose a hypothetical decision permits only doses of 0, 5, or 10 mg/day. Model predictions give benefit summaries $B(d)=0,30,45$ points and harm summaries $H(d)=0,4,12$ points at those respective doses. Use $w_B=1$ and $w_H=2$, in utility units per point.

For 0 mg/day, $U=0-2(0)=0$. For 5 mg/day, $U=30-2(4)=22$. For 10 mg/day, $U=45-2(12)=21$. Thus 5 mg/day maximizes utility among the permitted choices. The utility gain from 0 to 5 mg/day is 22 units, whereas the gain from 5 to 10 mg/day is $21-22=-1$ unit. There is no need to claim a continuous interior marginal equality: the decision is a direct comparison over the available dose set.

Diagnostics, Alternatives, and Limitations

A calculated optimum is only as credible as the dose-response summaries and utility specification behind it. Check that predictions are supported over the relevant dose range, that benefit and harm summaries use the intended population and time horizon, and that the weights have the stated units and interpretation. Plotting $U(d)$, $B(d)$, and $H(d)$ over the feasible interval can reveal implausible curvature, an optimum driven by extrapolation, or a boundary solution that was overlooked. Sensitivity analysis over defensible weights and model assumptions shows whether the recommended dose is stable or preference-sensitive.

The additive approach is useful when the decision maker is willing to exchange benefit and harm on a common utility scale. It differs from a benefit-only objective, which maximizes $B(d)$ and can ignore dose-related harm, and from a constrained objective, which may maximize benefit only among doses satisfying a prespecified harm requirement. Those objectives answer different decision questions; an additive optimum should not be presented as though it enforces a separate safety constraint.

The smooth marginal condition also has limits. A derivative may not exist for a discrete dose set or at a nonsmooth point; in those settings, compare allowable doses or use appropriate one-sided changes. Nonconcave utility can have several stationary points, so local optimization alone may select the wrong one. Finally, uncertainty in estimated dose-response curves can make the maximizing dose uncertain. The condition is a property of the specified utility curve; it does not, by itself, quantify uncertainty or validate the underlying model.

Common Mistakes

Equating total benefit and harm. The condition compares weighted marginal changes, $w_BB'(d)$ and $w_HH'(d)$, not $B(d)$ and $H(d)$.
Ignoring the weights. Equal raw slopes are generally not the condition when the weights differ. Compare slopes after applying the stated weights and units.
Treating every stationary dose as optimal. A stationary point may be a minimum, a local maximum, or one of several candidates. Check curvature, boundaries, and competing candidates.
Using a continuous derivative for a dose grid. When only specified dose levels are available, calculate utility at each permitted level.
Extrapolating to an unsupported optimum. A mathematical maximum outside the feasible or adequately modeled range is not a justified dose recommendation.

Key takeaway: For a differentiable additive utility, an interior optimum must balance weighted marginal benefit and weighted marginal harm. Establish feasibility and comparable units first, then verify that the candidate is a global maximum over the allowed doses.

Check Your Understanding

Answer each question, then compare with the solutions below.

For $U(d)=w_BB(d)-w_HH(d)$, state the interior marginal tradeoff condition and explain what each side represents.
If $U'(d)>0$ at every dose in a closed feasible interval, where is the maximum?
Suppose an interior stationary point has $U''(d^*)>0$. Does it satisfy the local maximum condition?
At three permitted dose levels, the utilities are 4, 9, and 7 units. Which dose should be selected, and why is a derivative condition unnecessary?

Solutions. (1) The condition is $w_BB'(d^*)=w_HH'(d^*)$: the left side is weighted marginal benefit per dose unit, and the right side is weighted marginal harm per dose unit. (2) Utility increases throughout the interval, so its maximum is at the upper boundary. (3) No. A positive second derivative indicates local curvature consistent with a minimum, not a maximum. (4) Select the dose with utility 9 units. The choices are discrete, so direct comparison identifies the maximum without a continuous derivative condition.