Nonlinear Utility for Benefit and Harm

Why Curvature Matters

A fixed increase in efficacy need not have the same value at every point on an efficacy scale. A change from little or no benefit to a moderate benefit may matter more to patients than the same numerical change near the maximum possible benefit. Conversely, a move to a more severe toxicity grade may impose a larger penalty than a move between mild grades. A linear utility curve cannot represent either pattern: it assigns the same utility change to every equal-sized change in its input.

A nonlinear curve makes the shape of these trade-offs explicit. Its curvature is a preference assumption, not a pharmacological law: observed treatment outcomes do not by themselves reveal how patients value those outcomes. The curve must be justified by preference information, a stated decision convention, or both.

Definition: A benefit utility function \(u_E(E)\) maps an efficacy outcome \(E\) to a utility value, with larger values representing greater benefit. A harm disutility function \(D(g)\) maps a harm or toxicity level \(g\) to a nonnegative penalty, with larger values representing worse harm. A benefit curve is concave if its marginal utility is nonincreasing; when it is twice differentiable, this means \(u_E''(E)\leq 0\). The condition \(u_E''(E)<0\) is sufficient for strict concavity. A disutility curve is convex if its marginal penalty is nondecreasing; when twice differentiable, this means \(D''(g)\geq 0\). The condition \(D''(g)>0\) is sufficient for strict convexity.

The derivative is the local change in utility per unit change in the clinical outcome. Thus \(u_E'(E)\) is measured in utility units per efficacy unit, and \(D'(g)\) in disutility units per toxicity unit. Concavity of benefit means successive equal increments in efficacy provide no larger marginal utility. Convexity of harm means successive equal increases in toxicity impose no smaller marginal disutility. These are distinct claims about the shape of the preference mapping.

Normalized Curves and Their Interpretation

A convenient diminishing-returns curve for an efficacy measure ranging from \(0\) to \(E_{\max}\) is the normalized exponential form

u_E(E)=\frac{1-\exp(-kE)}{1-\exp(-kE_{\max})},\qquad 0\leq E\leq E_{\max},\quad k>0.

The endpoints are fixed at \(u_E(0)=0\) and \(u_E(E_{\max})=1\). The parameter \(k\) has units inverse to those of \(E\). Differentiation gives

u_E'(E)=\frac{k\exp(-kE)}{1-\exp(-kE_{\max})}>0,\qquad u_E''(E)=-\frac{k^2\exp(-kE)}{1-\exp(-kE_{\max})}<0.

The curve is increasing and strictly concave: more efficacy always adds utility, but the marginal utility falls as efficacy rises. The value of \(k\) controls how quickly the curve bends; it should be calibrated to meaningful preference judgments rather than selected solely because the formula is convenient.

For toxicity on a severity scale from \(0\) to \(G_{\max}\), a simple increasing-penalty curve is

D(g)=\left(\frac{g}{G_{\max}}\right)^\gamma,\qquad 0\leq g\leq G_{\max},\quad \gamma>1.

It gives \(D(0)=0\) and \(D(G_{\max})=1\), and assigns relatively more of the total penalty to the more severe part of the scale. For \(g>0\),

D'(g)=\frac{\gamma g^{\gamma-1}}{G_{\max}^{\gamma}}\geq 0,\qquad D''(g)=\frac{\gamma(\gamma-1)g^{\gamma-2}}{G_{\max}^{\gamma}}>0.

At \(g=0\), \(D'(0)=0\) for \(\gamma>1\). The second derivative expression is stated for \(g>0\); for example, with \(\gamma=2\), the curve is twice differentiable at zero as well. The exponent is not a toxicity effect estimate. It encodes how the decision maker values severity increments on the chosen scale.

Worked Example: Calibrating Diminishing Returns for Efficacy

Suppose a hypothetical preference exercise uses an efficacy scale from 0 to 100 percentage points of improvement. The anchors assign utility 0 to no improvement and utility 1 to 100 points. A patient group judges 50 points to have utility 0.70. Fit the normalized exponential curve and interpret its values.

Set \(E_{\max}=100\) and impose \(u_E(50)=0.70\). With \(q=\exp(-50k)\), the anchor equation becomes

0.70=\frac{1-q}{1-q^2}=\frac{1}{1+q}, \qquad q=\frac{1}{0.70}-1=0.428571.

Therefore \(k=-\ln(q)/50=\ln(2.33333)/50\approx0.01695\) per percentage point. Substitution gives \(u_E(25)\approx0.4231\), \(u_E(50)=0.7000\), and \(u_E(75)\approx0.8813\), while \(u_E(100)=1.000\). The utility gain from 0 to 25 points is about 0.4231; from 75 to 100 it is only \(1.000-0.8813=0.1187\). The same 25-point efficacy increase has different value at different parts of this scale.

The fit reproduces the specified anchor by construction; that alone does not establish that the curve describes all preferences well. Additional preference judgments would be useful to assess its shape between and beyond the anchor points.

Fitting and Checking a Harm Curve

A power curve can be calibrated in the same anchor-based way. On an ordinal toxicity scale, the numerical coding itself is consequential: treating grades 0, 1, 2, 3, and 4 as equally spaced is a modeling assumption, not a property guaranteed by the grade labels. Define the scale and its endpoints before fitting, and check whether the elicited judgments support the assumed shape.

Worked Example: Fitting Increasing Disutility for Toxicity

Consider a hypothetical toxicity grade coded from 0 (none) through 4 (most severe). Suppose the preference anchors assign disutility 0 at grade 0, disutility 1 at grade 4, and disutility 0.25 at grade 2. Fit \(D(g)=(g/4)^\gamma\).

Substitute the grade-2 anchor:

0.25=\left(\frac{2}{4}\right)^\gamma=(0.5)^\gamma.

Since \(0.25=(0.5)^2\), the fitted exponent is \(\gamma=2\), giving \(D(g)=(g/4)^2\). The disutilities at grades 0, 1, 2, 3, and 4 are respectively 0, 0.0625, 0.2500, 0.5625, and 1.000. The successive one-grade increases in penalty are 0.0625, 0.1875, 0.3125, and 0.4375. Under this curve, the step from grade 3 to grade 4 carries seven times the disutility increase of the step from grade 0 to grade 1.

This calculation fits the given anchor exactly, but it does not validate equal spacing of the grades or the quadratic shape. If elicited values at other grades are available, compare them with the fitted values and investigate meaningful discrepancies.

The curvature statements follow directly from the definitions. If a twice-differentiable function has \(u_E''(E)\leq0\) throughout an interval, its derivative is nonincreasing there, so the marginal utility cannot rise as efficacy increases. For the power disutility curve with \(\gamma>1\), \(D''(g)>0\) for \(g>0\), so its derivative increases over positive severity levels. These derivative checks justify the diminishing-return and escalating-penalty interpretations; they do not establish that those shapes are appropriate for a particular patient population.

Using Both Curves in a Clinical Comparison

Separate curves can express nonlinear preferences for benefit and harm. To illustrate their implications, hold fixed a decision rule that subtracts a weighted disutility from efficacy utility:

U=u_E(E)-\lambda D(g),\qquad \lambda\geq0.

Here \(\lambda\) specifies the relative scale assigned to the harm penalty. This expression is useful for showing how curve shape affects a comparison, but the numerical result depends on the selected weight as well as on both curves. It is not justified merely by fitting each curve separately.

Worked Example: Comparing Two Hypothetical Benefit-Harm Profiles

Use the efficacy curve fitted above and the grade-squared disutility curve. Suppose the fixed hypothetical decision rule uses \(\lambda=0.50\). Treatment A has efficacy 50 and toxicity grade 2; treatment B has efficacy 75 and toxicity grade 3. Calculate their utilities on the specified scales.

For A, the fitted efficacy utility is \(u_E(50)=0.7000\), and \(D(2)=(2/4)^2=0.2500\). Thus

$$ U_A=0.7000-(0.50)(0.2500)=0.5750. $$

For B, \(u_E(75)\approx0.8813\) and \(D(3)=(3/4)^2=0.5625\), so

$$ U_B=0.8813-(0.50)(0.5625)=0.6001. $$

Under this specified rule, B exceeds A by approximately \(0.6001-0.5750=0.0251\) utility units. This is a conditional comparison, not a claim that B is clinically preferable under every patient’s preferences. A different harm weight or a different curvature could change the result.

Diagnostics, Alternatives, and Limitations

Fitting a nonlinear curve requires more than checking that the formula produces values between zero and one. Examine the underlying preference anchors, the endpoint definitions, and the behavior implied by the curve. Useful checks include:

Monotonicity: confirm that greater benefit does not reduce utility and that more severe harm does not reduce disutility, unless the decision context explicitly supports a different direction.
Anchor agreement: compare fitted values with every elicited point, not only the anchors used to estimate parameters. Plotting residual differences can reveal systematic misfit.
Local increments: calculate utility or disutility changes over clinically meaningful intervals. An extreme slope or unexpectedly large jump can expose a scale or parameter choice that deserves review.
Boundary behavior: verify values at the minimum and maximum and check that extrapolation beyond the elicited range is not being treated as reliable evidence.
Scale coding: for ordinal grades, test whether changing the numerical coding changes conclusions. Curvature depends on the chosen coding.

A linear function is the main alternative when equal-sized outcome changes are intended to have equal utility increments. A piecewise-linear curve can preserve that interpretation within ranges while allowing a slope change at a clinically justified threshold. Other smooth forms, including power or exponential functions, can represent curvature, but none is automatically preferable. Choose among them using preference evidence, interpretability, and fit, not convenience alone.

Nonlinear curves also have limits. A single curve may conceal substantial variation between patients; a group-level average curve need not describe any particular individual. Sparse elicitation data may not identify curvature reliably, and different functional forms can pass through the same few anchors while implying different values elsewhere. Finally, a score on a normalized scale is not itself a clinical outcome or a treatment effect. Its meaning depends on the anchors, the population whose preferences it represents, and any rule used to combine benefit and harm.

Common pitfall: “The last efficacy increment must be worth more because it is numerically the same size as the first, and toxicity grades are equally spaced by definition.” Both statements confuse the outcome scale with the preference scale. Equal numerical increments have equal utility only under a linear utility function; grade labels do not guarantee equal perceived severity gaps. Specify the coding, elicit or justify the curve, and inspect the implied increments before interpreting a comparison.

Key takeaway: Curvature determines how the value of an outcome changes across its range. Concave benefit utility encodes diminishing marginal benefit; convex harm disutility encodes increasing marginal penalty. Parameter estimates and clinical interpretations are credible only to the extent that the selected scales, preference anchors, and functional forms are defensible.

Check Your Understanding

Answer each question, then compare with the solutions below.

For \(u_E(E)=[1-\exp(-kE)]/[1-\exp(-kE_{\max})]\) with \(k>0\), what do the sign of the first derivative and the sign of the second derivative imply?
For \(D(g)=(g/4)^2\), calculate the disutility increment from grade 2 to grade 3.
Why does a grade-squared curve require more justification than simply observing that toxicity grades are numbered 0 through 4?
If an elicited anchor is \(u_E(50)=0.50\) on a 0-to-100 scale, what value of \(k\) makes the normalized exponential curve fit that anchor?

Solutions. (1) The first derivative is positive, so utility increases with efficacy; the second derivative is negative, so marginal utility decreases. (2) \(D(3)-D(2)=(3/4)^2-(2/4)^2=0.5625-0.2500=0.3125\) disutility units. (3) Numbering does not establish equal spacing in perceived severity; the squared curve and the grade coding are preference assumptions that should be checked against elicited judgments. (4) As in the calibration derivation, \(q=\exp(-50k)=1/0.50-1=1\), so \(k=-\ln(1)/50=0\). This is the limiting linear curve \(u_E(E)=E/100\); the strictly positive-\(k\) curved family approaches it as \(k\) approaches zero.