Why the Shape of a Utility Function Matters
A clinical cut point can represent an important change in how an outcome is judged: a response may be considered adequate above a target, or a biomarker may be considered acceptable below a safety limit. The rule used to translate outcomes into utility determines how differences around that cut point affect a dose comparison. A hard threshold treats outcomes on opposite sides of the cut point differently, even if they are very close. A hinge assigns a gradual, piecewise-linear transition, while a smooth function changes continuously.
These are not merely alternative mathematical descriptions of the same preference. They encode different judgments about whether small changes near the cut point matter, and about whether outcomes farther from it continue to earn additional value. The function should match the clinical decision, rather than being selected just because it is convenient to calculate.
The examples below use utilities between 0 and 1 and treat \(c\) as the midpoint of a transition for the hinge and smooth functions. These are particular, useful constructions—not the only possible utilities. If smaller outcomes are better, the direction must be reversed explicitly. A cut point and the transition parameters must also be justified clinically: the formulas do not determine what patients or decision-makers value.
Three Ways to Represent a Cut Point
A threshold utility records whether the cut point has been attained. For a larger-is-better outcome, define \(U_T(y)=1\) when \(y\geq c\), and \(U_T(y)=0\) otherwise. This convention assigns an outcome exactly equal to the cut point to the attained category. The function is discontinuous at \(c\): an arbitrarily small increase from just below \(c\) to \(c\) changes utility by one full unit.
A hinge utility replaces that jump with a clipped linear ramp. Let \(w>0\) be the half-width of the transition, in the outcome’s units. The utility is zero at and below \(c-w\), increases linearly between \(c-w\) and \(c+w\), and is one at and above \(c+w\):
This function is continuous, but its slope changes at the two knots \(c-w\) and \(c+w\). The slope in the transition region is \(1/(2w)\), so a narrower transition makes utility change more sharply per unit of outcome. The function does not reward further increases once \(y\geq c+w\), and it assigns the midpoint utility \(0.5\) at \(y=c\).
A smooth utility can use the logistic curve, with a positive scale parameter \(s\) measured in outcome units:
This function is continuous and differentiable, has utility \(0.5\) at \(c\), and approaches zero and one gradually in the lower and upper tails. The parameter \(s\) controls the rate of transition: a smaller \(s\) gives a steeper curve near \(c\), while a larger \(s\) gives a more gradual change. Unlike the clipped hinge, the logistic utility does not reach exactly zero or one at any finite outcome.
| Function | How it treats the cut point | What it retains or discards |
|---|---|---|
| Threshold | A jump from 0 to 1 at \(c\) | Retains attainment status; discards how far an outcome lies on either side |
| Hinge | A linear transition across \(c-w\) to \(c+w\) | Retains graded differences in the transition region; is flat outside it |
| Logistic smooth | A gradual transition centered at \(c\) | Retains graded differences throughout; approaches, but does not reach, its limiting values |
What the Transition Parameters Mean
The hinge width \(w\) and logistic scale \(s\) are not interchangeable. For the hinge, \(w\) sets the distance from the cut point to each end of the linear transition. For the logistic curve, \(s\) sets the scale of the gradual transition, without creating finite endpoints. Both parameters have the same units as \(y\), which makes their clinical interpretation more direct than a unitless tuning constant.
The smooth function is closely related to the threshold rule. For any fixed \(y<c\), \(U_S(y)\) approaches zero as \(s\) decreases toward zero; for any fixed \(y>c\), it approaches one. At \(y=c\), however, \(U_S(c)=0.5\) for every positive \(s\), whereas the threshold definition above gives \(U_T(c)=1\). Thus the logistic curve approaches the threshold away from the cut point, but does not converge to that exact threshold convention at the cut point itself.
This distinction matters when predicted outcomes cluster close to \(c\). A narrow smooth transition can behave much like a threshold for outcomes well away from the cut point, yet still distinguish outcomes near it. A hinge instead has a constant slope inside a specified interval and no slope outside. The choice of transition should be checked against the clinical meaning of small outcome differences, not chosen only to make one candidate dose look preferable.
Worked Example: Scoring Outcomes Around a Clinical Target
Suppose a hypothetical efficacy outcome is measured in response units, larger is better, and the clinical target is \(c=50\) response units. Compare the threshold rule with a hinge having \(w=10\) response units and a logistic utility having \(s=5\) response units.
For the hinge, the transition runs from \(c-w=40\) to \(c+w=60\). At \(y=45\), for example, \(U_H(45)=0.5+(45-50)/(2\times10)=0.25\). At \(y=55\), \(U_H(55)=0.5+(55-50)/20=0.75\).
For the smooth utility, at \(y=45\), \(U_S(45)=1/(1+\exp(-(-5/5)))=1/(1+\exp(1))\approx0.2689\). At \(y=55\), \(U_S(55)=1/(1+\exp(-1))\approx0.7311\).
| Outcome \(y\) (response units) | Threshold \(U_T(y)\) | Hinge \(U_H(y)\) | Smooth \(U_S(y)\) |
|---|---|---|---|
| 40 | 0 | 0 | 0.1192 |
| 45 | 0 | 0.25 | 0.2689 |
| 50 | 1 | 0.50 | 0.5000 |
| 55 | 1 | 0.75 | 0.7311 |
| 60 | 1 | 1 | 0.8808 |
At exactly 50 response units, the threshold convention gives utility 1, while the hinge and smooth utilities give 0.5. At 60, the hinge has reached its cap, but the smooth utility remains below 1. These differences follow from the definitions; they are not calculation errors. The selected function should reflect whether crossing the target is itself decisive, whether progress through a transition is valued, and whether very high outcomes should continue to gain value.
Using Utilities to Compare Predicted Dose Outcomes
For a dose \(d\), write \(Y_d\) for the outcome that would occur under that dose. When outcomes are uncertain, a common comparison is the expected utility \(E[U(Y_d)]\): apply the chosen utility to each possible outcome, then average using the outcome probabilities. This calculation requires a specified outcome distribution or a set of predicted scenarios, as well as a utility function whose direction and parameters are appropriate to the decision.
The three functions can rank doses differently because they score outcomes differently. The threshold function values the chance of attaining the cut point, while the hinge and smooth functions also distinguish some outcomes below or above it. The hinge’s flat regions can still make distinct outcomes equivalent, and the logistic curve gives nonzero weight to outcomes on either side.
Worked Example: A Dose Ranking Changes With the Utility Shape
Consider two hypothetical dose predictions for a larger-is-better outcome with \(c=50\) response units. Dose A produces 52 response units with probability 1. Dose B produces 40 response units with probability 0.5 and 70 response units with probability 0.5. Use the hinge with \(w=10\) and the logistic function first with \(s=5\).
For the threshold function, Dose A always attains the target, so \(E[U_T(Y_A)]=1\). Dose B attains it only in the 70-unit scenario, so \(E[U_T(Y_B)]=0.5(0)+0.5(1)=0.5\). The threshold therefore ranks A above B.
For the hinge, Dose A has utility \(U_H(52)=0.5+(52-50)/20=0.60\). Dose B has utility 0 at 40 and 1 at 70, so \(E[U_H(Y_B)]=0.5(0)+0.5(1)=0.50\). The hinge also ranks A above B.
For the logistic function with \(s=5\), Dose A has \(U_S(52)=1/(1+\exp(-2/5))\approx0.5987\). For Dose B, the utilities are \(U_S(40)=1/(1+\exp(2))\approx0.1192\) and \(U_S(70)=1/(1+\exp(-4))\approx0.9820\). Therefore, \(E[U_S(Y_B)]=0.5(0.1192)+0.5(0.9820)=0.5506\). Dose A ranks higher, with expected utility about 0.5987 compared with 0.5506 for Dose B.
Now use the same logistic definition with \(s=15\). Dose A has \(U_S(52)=1/(1+\exp(-2/15))\approx0.5333\). Dose B has \(U_S(40)=1/(1+\exp(10/15))\approx0.3392\) and \(U_S(70)=1/(1+\exp(-20/15))\approx0.7914\), giving \(E[U_S(Y_B)]=0.5(0.3392)+0.5(0.7914)=0.5653\). Under this broader transition, B ranks higher than A. The reversal is a consequence of changing the utility scale parameter, not a change in the predicted outcomes. It shows why a dose ranking should be accompanied by a justification and sensitivity assessment for the utility parameters.
Changing the Outcome Direction
The formulas above assume larger outcomes are better. For a continuous adverse outcome where smaller is better, define the utility to decrease as the outcome rises. One direct approach is to replace \(y-c\) by \(c-y\) in the hinge and logistic formulas, and define threshold attainment as \(y\leq c\). This keeps the cut point in the original outcome units and makes the direction of preference explicit.
Worked Example: A Maximum Acceptable Adverse Biomarker
Suppose a hypothetical continuous adverse biomarker is measured in fold-of-reference units, with smaller values preferred and a maximum acceptable cut point \(c=3\). For a hinge, use \(w=1\); for a logistic utility, use \(s=0.5\). The threshold rule is \(U_T(y)=1\) for \(y\leq3\), and 0 otherwise.
For the decreasing hinge, utility is 1 at and below \(c-w=2\), declines linearly between 2 and 4, and is 0 at and above 4. At \(y=2.5\), \(U_H(2.5)=0.5+(3-2.5)/(2\times1)=0.75\). At \(y=3.5\), \(U_H(3.5)=0.5+(3-3.5)/2=0.25\).
For the decreasing logistic function, \(U_S(y)=1/(1+\exp(-(c-y)/s))\). At \(y=2.5\), \(U_S(2.5)=1/(1+\exp(-1))\approx0.7311\); at \(y=3.5\), \(U_S(3.5)=1/(1+\exp(1))\approx0.2689\). At \(y=3\), the threshold utility is 1, while the hinge and logistic utilities are each 0.5. The calculation confirms that the continuous utilities reward lower biomarker values and that the threshold uses the stated \(y\leq c\) convention.
Assumptions, Diagnostics, and Failure Modes
Before applying one of these functions to dose predictions, check the following conditions. The outcome must be defined on a scale where the cut point has a meaningful clinical interpretation. The direction of preference must be correct throughout the range being scored. The chosen function and its parameters must represent the decision-maker’s intended tradeoffs, and predicted outcomes must be expressed in the same units as the cut point and transition parameters. If the function is part of a larger utility, its relative scale must also be compatible with the other components.
Plot or tabulate \(U(y)\) over the clinically plausible outcome range. This simple diagnostic can expose a reversed direction, an unintended jump, an overly narrow transition, or a hinge that becomes flat where further benefit or harm should still matter. Also inspect predictions near the cut point: if a small modeling or measurement difference moves many predictions across a threshold, a threshold-based dose ranking may be especially sensitive to that classification. For hinge and smooth functions, compare plausible values of \(w\) or \(s\); a ranking that changes under modest, clinically plausible choices is not robust to preference specification.
- Using a threshold when distance matters. The threshold gives the same utility to every outcome above \(c\), and the same utility to every outcome below \(c\). If the extent of benefit or shortfall matters, that loss of information may be inappropriate.
- Treating a hinge as smooth. The hinge is continuous but has slope changes at its knots. Optimization or interpretation that assumes a continuously differentiable function can therefore be misleading at those points.
- Ignoring units. A width of 5 has no clinical meaning without specifying whether the outcome is measured in response units, concentration units, or another scale. A change of measurement units requires corresponding changes to \(c\), \(w\), or \(s\).
- Reversing the outcome direction. Applying a larger-is-better formula to an adverse measure can reward more harm. State explicitly whether larger or smaller outcomes are preferred.
- Assuming the utility parameters are model parameters. The cut point and transition width or scale describe the preference mapping. They are not estimated outcome-model parameters merely because they appear in a dose analysis.
The main alternatives answer different questions. A threshold is appropriate when the decision genuinely depends on whether a target is met and magnitude beyond attainment is irrelevant. A hinge is useful when preferences support a bounded linear transition with clear endpoints. A logistic smooth function is useful when a gradual, differentiable transition is desired and small differences remain valued in the tails. None is universally superior, and a smooth curve is not automatically more clinically realistic merely because it has no corners.
A Short Justification for the Smooth Approximation
For \(y\ne c\), the logistic utility can be written in terms of \((y-c)/s\). If \(y>c\), then \((y-c)/s\) grows without bound as \(s\) approaches zero, so \(\exp(-(y-c)/s)\) approaches zero and \(U_S(y)\) approaches 1. If \(y<c\), then \(-(y-c)/s\) grows without bound, so the denominator grows without bound and \(U_S(y)\) approaches 0. This is why reducing \(s\) makes the smooth curve approximate a threshold away from the cut point. At \(y=c\), the exponent is zero for every \(s>0\), giving \(U_S(c)=1/2\), so the equality convention remains different.
The derivative also clarifies the role of \(s\). Writing \(U_S(y)=1/(1+\exp(-(y-c)/s))\), differentiation gives \(U'_S(y)=U_S(y)(1-U_S(y))/s\). The slope is positive, largest at \(y=c\), where it equals \(1/(4s)\), and tends toward zero in both tails. Thus increasing \(s\) lowers the maximum slope and spreads the transition over a wider outcome range.
Check Your Understanding
Answer each question, then compare your reasoning with the solutions.
- For a larger-is-better outcome with \(c=100\), what threshold utility is assigned to \(y=99.9\) and \(y=100\)? What does this illustrate?
- A hinge has \(c=20\) and \(w=4\). Calculate its utility at \(y=18\), \(y=20\), and \(y=24\).
- For a logistic utility with \(c=10\), explain qualitatively what happens to the transition when \(s\) increases.
- A lower-is-better adverse outcome has cut point \(c=5\). Write the threshold attainment rule and state the utility at \(y=4.8\) and \(y=5.2\).
- Two doses have the same probability of exceeding a target, but one has much higher outcomes when it exceeds the target. Which utility form may distinguish them, and what additional preference assumption does that require?
Solutions. 1. \(U_T(99.9)=0\) and \(U_T(100)=1\). An arbitrarily small change across the cut point causes a full one-unit jump in utility. 2. The hinge transition runs from \(20-4=16\) to \(20+4=24\). At 18, \(U_H(18)=0.5+(18-20)/8=0.25\); at 20, it is 0.5; at 24, it is 1. 3. Increasing \(s\) makes the logistic transition more gradual: the slope at \(c\) is \(1/(4s)\), so it decreases as \(s\) increases. 4. Attainment is \(y\leq5\). Thus the threshold utility is 1 at 4.8 and 0 at 5.2. 5. A hinge or smooth utility may distinguish the magnitudes, whereas the threshold treats all values above the target equally. This requires a preference that additional outcome magnitude has value, represented by the chosen transition and its parameters.