Averaging Utilities Over Outcome Distributions

Why Average Utility Across Outcomes?

A treatment does not usually produce the same clinical outcome for every patient. Even if its average outcome is known, patients may experience several possible benefit and harm states, each with its own value to patients or decision makers. To compare treatments using utility, the relevant quantity is generally the average of the utilities of the possible outcomes, weighted by how likely those outcomes are—not the utility assigned to a single average outcome.

This distinction matters whenever the outcome distribution contains meaningful variation and the utility assigned to an outcome does not preserve averages. It also matters when the mean is not itself a possible clinical outcome. The calculation is straightforward once the outcome states, their probabilities, and their utilities are specified. The difficult work is ensuring they refer to the same clinical context and preference profile.

Definition: Let $T=t$ denote a treatment and $Y$ its clinical outcome, with treatment-specific outcome distribution $P(Y=y\mid T=t)$. For a specified utility function $u(y)$, the expected utility of treatment $t$ is the average utility over that distribution: $$ EU(t)=E[u(Y)\mid T=t]. $$ For a discrete outcome with possible values $y_1,\ldots,y_K$, this is $$ EU(t)=\sum_{k=1}^{K}P(Y=y_k\mid T=t)u(y_k). $$ For a continuous outcome with density $f_t(y)$, it is $$ EU(t)=\int u(y)f_t(y)\,dy. $$

Utilities must be on a common scale across treatments. In the examples below, they are expressed in utility points, with larger values representing more preferred outcomes. The distribution must describe the outcomes relevant to the decision, and the outcome categories must cover the possibilities being considered. For discrete outcomes, each category is mutually exclusive, its probability is nonnegative, and the probabilities sum to one. For continuous outcomes, the density integrates to one. The utility function and the distribution must refer to the same outcome definition and time horizon.

These calculations assume a specified preference profile and treat the outcome probabilities and utilities as inputs. If preferences differ between patients, a single utility function may not represent everyone; modeling that heterogeneity requires care, as discussed in the earlier material on modeling heterogeneity in benefit-risk preferences. Here, hold the preference profile fixed so the effect of averaging over outcomes is clear.

Discrete Outcome Distributions

For a discrete outcome, expected utility is a probability-weighted sum. The probability weights describe how often each outcome is expected under a treatment; the utility values describe how desirable those outcomes are on the chosen scale. These are different inputs and should not be confused. A less desirable outcome can still contribute substantially to expected utility if its probability is large.

Worked Example: Comparing Treatments Across Benefit-Risk States

Hypothetical data. Consider three mutually exclusive clinical states at a specified follow-up: substantial benefit without a limiting adverse event, partial benefit without a limiting adverse event, and no benefit with a limiting adverse event. A decision maker assigns utilities of 100, 70, and 20 utility points to these states, respectively. Treatment A has probabilities 0.45, 0.40, and 0.15; treatment B has probabilities 0.55, 0.30, and 0.15. Calculate and compare expected utility.

First check that each distribution is complete: for A, $0.45+0.40+0.15=1.00$; for B, $0.55+0.30+0.15=1.00$. All probabilities are nonnegative, and the same outcome definitions and utilities apply to both treatments.

For A, $$ EU(A)=0.45(100)+0.40(70)+0.15(20) =45+28+3=76.0\text{ utility points}. $$

For B, $$ EU(B)=0.55(100)+0.30(70)+0.15(20) =55+21+3=79.0\text{ utility points}. $$

On this specified utility scale and under these hypothetical distributions, B has 3.0 more expected utility points than A. This comparison does not show that B is better for every patient, nor does it establish a treatment effect in a real population. It summarizes the stated distributions using the stated utilities.

The calculation uses every outcome category, including one with an unchanged probability across treatments. Omitting a category is valid only if its contribution is handled consistently—for example, by explicitly defining utilities relative to a reference state and retaining the corresponding probability-weighted values. Simply dropping a category because it appears less important can change the comparison.

Why the Distribution Cannot Usually Be Replaced by Its Mean

For a discrete outcome, let the treatment-specific probabilities be $p_k=P(Y=y_k\mid T=t)$, so $\sum_k p_k=1$. The mean outcome is $\mu_t=\sum_kp_ky_k$. Expected utility is $\sum_kp_ku(y_k)$, whereas applying utility to the mean gives $u(\mu_t)$. These expressions are not generally the same: the first averages utility values at the possible outcomes, and the second evaluates utility at one summary value.

Key result: The distribution-based calculation is $E[u(Y)\mid T=t]$. Replacing it with $u(E[Y\mid T=t])$ is justified when the two quantities are equal—for example, when $u$ is affine over the relevant outcome range, or when the outcome is constant. In other cases, equality may occur by coincidence, but it must be checked rather than assumed.

A short algebraic comparison makes the issue explicit. Since $\mu_t=\sum_kp_ky_k$, the difference between expected utility and utility at the mean is $$ E[u(Y)\mid T=t]-u(\mu_t) =\sum_kp_k\{u(y_k)-u(\mu_t)\}. $$ There is no general reason for the weighted terms on the right to sum to zero. If $u(y)=a+by$ across the relevant range, then $$ \sum_kp_k u(y_k) =\sum_kp_k(a+by_k) =a\sum_kp_k+b\sum_kp_ky_k =a+b\mu_t =u(\mu_t), $$ because $\sum_kp_k=1$. This establishes why the shortcut works for an affine utility function. Without that condition or another reason for equality, calculate from the distribution.

Worked Example: Same Mean Outcome, Different Expected Utility

Hypothetical data. A clinical response is classified as low, moderate, or high, with outcome scores 0, 1, and 2. A previously specified utility table assigns 0, 40, and 100 utility points to those outcomes. Treatment A yields low response with probability 0.50 and high response with probability 0.50. Treatment B yields moderate response with probability 1.00.

Both treatments have the same mean response score: $$ E[Y\mid A]=0.50(0)+0.50(2)=1,\qquad E[Y\mid B]=1.00(1)=1. $$ Applying utility only to that mean gives $u(1)=40$ points for both treatments.

The distribution-based calculations instead give $$ EU(A)=0.50(0)+0.50(100)=50.0\text{ utility points}, $$ and $$ EU(B)=1.00(40)=40.0\text{ utility points}. $$

Thus the two treatments have equal mean response scores but different expected utilities under the supplied utility table: A exceeds B by 10.0 utility points. This result follows from the outcome distributions and assigned utilities; the mean score alone does not retain enough information to reproduce it.

Continuous Outcomes and the Mean-Outcome Shortcut

For a continuous outcome, the same principle applies with an integral. The density $f_t(y)$ weights the utility $u(y)$ at each possible outcome value. If the utility function is affine, integrating the utility across the distribution gives the utility of the mean. The example below shows this special case and checks the calculation directly.

Worked Example: A Continuous Outcome With Linear Utility

Hypothetical data. Let $Y$ be symptom improvement in points, with $Y\sim\text{Uniform}(0,10)$ under treatment A. Under treatment B, improvement is always 5 points. The supplied utility function is $u(y)=10+8y$, measured in utility points. Assume the distributions and utility function refer to the same follow-up and preference profile.

Under A, the density is $f_A(y)=0.1$ for $0\le y\le10$, and the mean improvement is 5 points. Directly averaging utility over the distribution gives $$ EU(A)=\int_0^{10}(10+8y)(0.1)\,dy =0.1[10y+4y^2]_0^{10} =0.1(100+400) =50.0\text{ utility points}. $$

Under B, the outcome is fixed at 5 points, so $$ EU(B)=u(5)=10+8(5)=50.0\text{ utility points}. $$

The two distributions differ, but their expected utilities are equal in this example because the utility function is affine and both mean improvements are 5 points. This equality should not be generalized to a different utility function without checking it.

Practical Checks, Alternatives, and Limitations

Before interpreting an expected-utility comparison, check the construction of both the distribution and the utility inputs. These diagnostics do not prove that the inputs are clinically appropriate, but they help detect errors that can make the arithmetic misleading.

Check probability totals. Confirm that discrete outcome probabilities are nonnegative and sum to one, or that the continuous density integrates to one. If categories omit possible outcomes, state how those outcomes are handled.
Check outcome definitions. Categories must be mutually exclusive and collectively cover the outcome space being summarized. Keep the time horizon and clinical definitions consistent across treatments.
Check the utility scale. Apply the same utility values, reference, and units to corresponding outcomes under each treatment. Do not average utilities from incompatible scales.
Check the calculation. Each contribution is probability multiplied by the utility of that specific outcome. A plausible-looking total can still be wrong if a probability is paired with the wrong utility.
Check whether the mean is informative enough. If utility is not affine over the relevant range, or if outcome profiles differ despite having the same mean, report distribution-based expected utility rather than relying on the mean alone.

The main alternative is to compare treatments using a mean clinical endpoint. That can be useful when the endpoint itself is the decision target, but it does not by itself summarize the preference-weighted value of all possible outcomes. Another shortcut, applying utility to the mean endpoint, is appropriate only when its equality to expected utility is established. For discrete outcomes, the probability-weighted sum is usually transparent and easy to audit; for continuous outcomes, numerical integration may be needed if the integral has no convenient analytic form.

Expected utility is a summary, not a complete description of a treatment. Different distributions can have the same expected utility, and a higher expected utility does not mean every patient benefits. Results depend on the chosen outcome space, the utility assignments, and the probability distributions. If any of these inputs are estimated imprecisely or are uncertain, a single calculation can hide that uncertainty; this tutorial treats them as specified inputs rather than quantifying their uncertainty.

A common mistake is to calculate the mean outcome first and then assume that its utility represents the treatment’s average value. The reasoning is that “the mean captures the outcome, so the utility of the mean must capture the utility.” The correction is to distinguish $E[u(Y)\mid T=t]$ from $u(E[Y\mid T=t])$, calculate the former from the distribution, and use the shortcut only when equality is justified.

Check Your Understanding

Try each problem, then compare your work with the solutions below.

A treatment has two possible outcomes with utilities 30 and 90 points and probabilities 0.25 and 0.75. Calculate expected utility.
Two treatments have identical mean outcome values. Is their expected utility necessarily identical? State the condition that would make equality with utility at the mean hold across the relevant range.
A discrete outcome distribution has probabilities 0.40, 0.35, and 0.20. What should be checked before calculating expected utility, and what is the total of the listed probabilities?
For a continuous outcome, $Y\sim\text{Uniform}(0,4)$ and $u(y)=5+2y$. Calculate expected utility by integration and compare it with utility at the mean.

Solutions. (1) $EU=0.25(30)+0.75(90)=7.5+67.5=75.0$ utility points. (2) No. Different distributions with the same mean can have different expected utilities; an affine utility function over the relevant range ensures expected utility equals utility at the mean. (3) Check that probabilities are nonnegative and sum to one, and verify that the outcome categories cover the relevant possibilities. The listed probabilities total $0.40+0.35+0.20=0.95$, so the distribution is incomplete as stated; the missing probability of 0.05 must be accounted for before calculating. (4) The density is $1/4$ on $[0,4]$, so $EU=\int_0^4(5+2y)(1/4)\,dy=(1/4)[5y+y^2]_0^4=(1/4)(20+16)=9.0$ utility points. The mean outcome is 2, and $u(2)=5+2(2)=9.0$ points, equal here because the utility function is affine.