Incorporating Treatment Costs Into Patient Utility

Why Include Costs in Patient Utility?

An allocation target can account for more than treatment response and treatment-related harm. Participants may also face travel, time away from work, inconvenience, or out-of-pocket expenses. If these differ between arms, they can affect the expected utility of enrolling a participant in each arm and therefore change an ethically motivated allocation target.

The key is to define what “cost” means in the objective. A dollar is not automatically a patient-utility point. Monetary expenses must either be reported separately as a distinct objective or translated into utility using an explicit valuation. Likewise, a visit burden already recorded in utility should not be counted again as a monetary cost unless the added amount represents a separate burden.

Definition: Let $B_a$ denote expected treatment benefit in arm $a$, $H_a$ expected treatment-related harm, $D_a$ expected nonmonetary participant burden, and $C_a$ expected monetary cost per participant, for $a\in\{C,T\}$, where $C$ denotes control and $T$ treatment. For a stated conversion value $\lambda$, measured in utility points per dollar, define arm-specific net utility as $U_a=B_a-H_a-D_a-\lambda C_a$. The treatment allocation share is $q$, so the control share is $1-q$.

This formulation assumes that all terms in the equation have been put on a compatible utility scale. For example, $B_a$, $H_a$, and $D_a$ might be measured in utility points, while $C_a$ is in dollars and $\lambda$ is utility points per dollar. Then $\lambda C_a$ is in utility points. The value of $\lambda$ is a substantive valuation choice, not a statistical quantity that can be inferred from the allocation formula alone.

From Arm Utilities to an Allocation Target

Suppose a trial has a feasible set of treatment shares $q\in[L,U]$, where $0\leq L\leq U\leq1$. The bounds may encode an ethical requirement to retain a minimum control group or another allocation constraint. Assume that each enrolled participant contributes the expected utility of the arm to which that participant is assigned, and that the arm-specific expected utilities do not change with $q$. The expected utility per participant under share $q$ is

$$ W(q)=qU_T+(1-q)U_C=U_C+q(U_T-U_C). $$

Write $\Delta U=U_T-U_C$. This difference can be expanded to show exactly how costs and other valuations affect the target:

\Delta U=(B_T-B_C)-(H_T-H_C)-(D_T-D_C)-\lambda(C_T-C_C).

The comparison depends on differences between arms. A cost incurred equally in both arms lowers both arm utilities but does not change $\Delta U$, so it does not alter the welfare-maximizing share under this model. An additional treatment-arm cost lowers $\Delta U$ when $\lambda>0$; a lower treatment-arm cost can raise it.

The allocation result follows from a short derivation. For any feasible shares $q_1$ and $q_2$,

W(q_2)-W(q_1)=(q_2-q_1)\Delta U.

If $\Delta U>0$, increasing $q$ increases expected utility, so the maximizing feasible share is $q=U$. If $\Delta U<0$, increasing $q$ decreases expected utility, so the maximizing feasible share is $q=L$. If $\Delta U=0$, then $W(q)=U_C$ for every feasible $q$, and all shares in $[L,U]$ tie on this welfare objective.

Result: Under fixed arm-specific expected utilities and a feasible interval $q\in[L,U]$, the utility-maximizing target is $U$ when $U_T>U_C$, $L$ when $U_T<U_C$, and any feasible share when $U_T=U_C$. This is a target for expected welfare under the specified valuation; it is not a claim that every individual would prefer the arm with the higher average utility.

Worked Example: Adding Monetary Costs to Arm Utilities

Consider a hypothetical trial with a feasible treatment-share range of $0.25$ to $0.75$. Suppose the benefit, harm, participant burden, and monetary cost per participant are as follows. The chosen valuation is $\lambda=0.02$ utility points per dollar.

Arm	Benefit (points)	Harm (points)	Nonmonetary burden (points)	Cost (dollars)
Control	4.8	0.1	0.4	$10
Treatment	6.2	0.3	1.0	$35

Control utility. The cost term is $0.02(10)=0.20$ utility points, so

$$ U_C=4.8-0.1-0.4-0.02(10)=4.10. $$

Treatment utility. The cost term is $0.02(35)=0.70$ utility points, so

$$ U_T=6.2-0.3-1.0-0.02(35)=4.20. $$

Thus, $\Delta U=4.20-4.10=0.10$ utility points per participant. Since this difference is positive, expected utility increases with the treatment share. The maximum on the feasible interval is therefore $q=0.75$. At that share,

$$ W(0.75)=0.75(4.20)+0.25(4.10)=3.15+1.025=4.175. $$

For comparison, at the lower bound, $W(0.25)=0.25(4.20)+0.75(4.10)=1.05+3.075=4.125$. Under these stated utilities and constraints, the upper-bound target has higher expected utility by $4.175-4.125=0.050$ points per participant.

Break-Even Cost: Count Every Burden Once

A break-even calculation asks how much additional monetary cost would make the arms equal in utility, given the other terms already included. The calculation must start with the full current utility difference. In particular, it must include any existing participant burden, such as extra visits, before solving for the additional cost.

Worked Example: Finding the Break-Even Incremental Cost

In a hypothetical comparison, treatment provides 6 benefit points and control provides 4. Treatment has 0.4 harm points compared with 0.2 for control. Treatment also has a visit burden of 1.3 points compared with 0.1 for control. Let $x$ be an additional treatment-only monetary cost in dollars, beyond any costs already represented, and let $\lambda=0.003$ utility points per dollar. Suppose feasible treatment shares are $0.10$ to $0.80$.

Before the added monetary cost, the treatment-minus-control utility difference is benefit difference minus harm difference minus visit-burden difference:

$$ (6-4)-(0.4-0.2)-(1.3-0.1)=2-0.2-1.2=0.6. $$

The additional cost contributes $0.003x$ points of disutility, giving

\Delta U=0.6-0.003x.

At break-even, set $\Delta U=0$. Solving gives $0.003x=0.6$, so $x=0.6/0.003=\$200$. If the additional cost is $100, then $\Delta U=0.6-0.003(100)=0.30$, and the welfare-maximizing feasible share is $q=0.80$. At $200, $\Delta U=0$, so all shares from $0.10$ to $0.80$ tie on expected welfare. At $250, $\Delta U=0.6-0.003(250)=-0.15$, and the maximizing feasible share is $q=0.10$.

The visit burden is essential to this calculation. Omitting it would incorrectly use $2-0.2=1.8$ as the advantage before additional cost and yield a break-even amount of $600. But at $600, the full difference is $2-0.2-1.2-0.003(600)=-1.2$, not zero. The correct break-even amount is $200 because the already-counted visit burden reduces the initial treatment advantage from 1.8 to 0.6 points.

Valuation Choices and Sensitivity

The conversion value $\lambda$ determines how strongly a dollar difference affects utility. It should be stated, justified, and examined for sensitivity. It can reflect a specified preference or decision rule; it should not be presented as an objective scientific constant. If stakeholders instead want to keep dollars and health utility separate, they need a multi-objective decision rule rather than silently treating one dollar as a fixed number of utility points.

Worked Example: How the Dollar Valuation Changes the Target

Suppose the treatment-control difference in benefit is 0.8 points, the difference in harm is 0.1 points, and the difference in nonmonetary burden is 0.2 points. Treatment costs an additional $30 per participant. The feasible treatment share is $q\in[0.30,0.70]$. Thus,

\Delta U=0.8-0.1-0.2-30\lambda=0.5-30\lambda.

If $\lambda=0.01$ points per dollar, then $\Delta U=0.5-30(0.01)=0.20$ points, so the maximizing target is $q=0.70$. If $\lambda=0.02$, then $\Delta U=0.5-30(0.02)=-0.10$, so the maximizing target is $q=0.30$. The break-even valuation satisfies $0.5-30\lambda=0$, hence $\lambda=0.5/30=0.01667$ points per dollar. At this value, all feasible shares tie under the stated objective.

This example does not show that one valuation is correct. It shows that the allocation conclusion depends on the valuation when the net arm utilities are close. Reporting the assumed conversion and the resulting sensitivity makes that dependence visible.

Checks, Comparisons, and Limitations

Before calculating a target, check the components and assumptions of the utility model. The same person’s travel time should not be entered once as a burden in $D_a$ and again as a dollar-valued cost in $C_a$, unless those terms measure distinct consequences. Confirm that costs are per participant, use a stated time horizon, and refer to the same perspective in both arms. A participant’s copayment, a health system’s expenditure, and a societal cost are different quantities; the chosen perspective must match the decision being analyzed.

The linear allocation result also relies on arm utilities being fixed as the share changes. If costs or benefits depend on how many participants receive an arm, the utility function need not be linear, and the endpoint rule need not apply. If the quantities are estimates rather than known values, uncertainty in them can make the selected endpoint fragile. Sensitivity calculations such as the previous example reveal when modest valuation changes reverse the direction of the target, but do not remove that uncertainty.

This welfare objective differs from a precision objective. The Neyman allocation in earlier material uses outcome variances and the contrast of interest to target statistical precision. The utility-based target here uses expected arm utilities and ethical constraints. A target that maximizes one objective does not automatically maximize the other. The feasible interval matters as well: when treatment has higher utility, the result is its largest allowed share, not necessarily allocation of everyone to treatment.

Useful diagnostics include checking the sign of each component in $\Delta U$, recomputing the difference from the arm totals, and evaluating $W(q)$ at both feasible endpoints. If $\Delta U>0$, the computed welfare at $U$ must exceed that at $L$ by $(U-L)\Delta U$; if not, an arithmetic or sign error has occurred. If a small change in $\lambda$ changes the sign of $\Delta U$, report that the allocation recommendation is valuation-sensitive rather than implying a robust preference.

Double-counting: Subtracting a visit burden in $D_a$ and also translating the same visit into $C_a$ counts it twice.
Inconsistent units: Subtracting dollars directly from utility points is invalid without a stated conversion.
Ignoring common costs: A common cost changes absolute utility but, under this model, does not change the arm difference or target.
Overstating the conclusion: A maximizing share is optimal only for the stated utility values, conversion, constraints, and assumptions; it is not a universal ethical recommendation.

Key takeaway: Calculate each arm’s net utility using explicit, nonduplicated terms; form $\Delta U=U_T-U_C$; and choose the appropriate endpoint of the feasible treatment-share interval. A break-even cost must be solved from the complete utility difference, including burdens already counted.

Check Your Understanding

Answer each question, then compare your reasoning with the solutions.

Control has utility $3.6$, treatment has utility $3.9$, and feasible treatment shares are $0.20$ to $0.65$. Which share maximizes expected utility?
A treatment arm costs an additional $40 per participant. If $\lambda=0.015$ utility points per dollar, how many utility points does this cost difference contribute to treatment’s utility difference?
Benefit favors treatment by 1.1 points, treatment has 0.3 additional harm points and 0.4 additional burden points, and its additional cost is $20. With $\lambda=0.01$, calculate $\Delta U$ and identify the maximizing share if $q\in[0.25,0.75]$.
Why can a cost shared equally by both arms fail to change the welfare-maximizing treatment share?
A break-even calculation omits a treatment-specific visit burden already included in the utility model. What is the error, and how should the calculation be corrected?

Solutions. (1) $\Delta U=3.9-3.6=0.3>0$, so the maximizing share is the upper bound, $q=0.65$. (2) The cost contribution is $-\lambda(40)=-0.015(40)=-0.60$ utility points to the treatment-control difference. (3) $\Delta U=1.1-0.3-0.4-0.01(20)=1.1-0.3-0.4-0.2=0.20$ points, so choose $q=0.75$. (4) The same cost is subtracted from each arm utility, so it cancels in $U_T-U_C$; with fixed utilities and the stated objective, it changes neither the sign of the difference nor the target. (5) Omitting the burden overstates treatment’s advantage and makes the calculated break-even cost too large. Include the burden in the initial arm difference, then solve $\Delta U=0$ for the additional monetary cost.