Bayesian Expected Utility for Benefit-Risk

From Posterior Uncertainty to a Benefit-Risk Decision

A benefit-risk analysis often begins with uncertain quantities: response probabilities, adverse-event risks, or other treatment outcomes. A utility model translates those quantities into a common scale for evaluating the trade-off. Bayesian expected utility combines the two: it averages the utility implied by the treatment parameters over their posterior distribution, conditional on the observed data.

The result is useful for decisions, but it answers a specific question. It describes expected utility under a specified statistical model and a specified utility function. It does not establish that the utility function represents every patient's preferences, nor does it by itself establish that a treatment causes the outcomes represented in the model.

Definition: Let $D$ denote the observed data, $\theta_t$ the uncertain efficacy and safety parameters for treatment $t$, and $u_t(\theta_t)$ the treatment's expected utility conditional on those parameters. The posterior expected utility is $$ EU_t(D)=E\{u_t(\theta_t)\mid D\} =\int u_t(\theta_t)\,p(\theta_t\mid D)\,d\theta_t. $$ The utility function, including its weights and scale, is held fixed in this definition. The expectation averages over posterior parameter uncertainty.

Here, $u_t(\theta_t)$ may already average over possible patient outcomes conditional on the parameters. For example, if response has utility $v_R$, no response has utility $v_N$, and the response probability is $p_R$, then the conditional expected response utility is $p_Rv_R+(1-p_R)v_N$. Additional terms can represent harms or other attributes, provided the utility model specifies how they combine.

If treatment $A$ and treatment $B$ are compared using the same utility scale, define their incremental utility at a parameter draw as $\Delta(\theta)=u_A(\theta)-u_B(\theta)$. Two different summaries are useful:

E\{\Delta(\theta)\mid D\}=EU_A(D)-EU_B(D), \qquad P\{\Delta(\theta)>0\mid D\}.

The first is the posterior mean incremental utility, measured in utility units. The second is the posterior probability that $A$ has greater utility than $B$ under a draw from the posterior and the specified utility model. Neither quantity is the probability that a treatment is “truly best” independently of the model, utility specification, and target population.

What Is Being Averaged?

The averaging has two layers when utility is defined on individual outcomes. For a parameter value $\theta$, first average utility over the outcome distribution implied by that parameter. Then average that conditional expected utility over the posterior distribution of $\theta$. For a discrete outcome $Y$, the identity is

E\{u(Y)\mid D\} =\int \left[\sum_y u(y)P(Y=y\mid\theta)\right]p(\theta\mid D)\,d\theta.

This follows by conditioning on $\theta$: the inner sum is $E\{u(Y)\mid\theta\}$, and averaging that quantity over $p(\theta\mid D)$ gives the posterior expectation of outcome utility. It is important not to replace this calculation with utility evaluated at posterior mean parameters unless the utility expression is linear in those parameters. In general,

E\{u(\theta)\mid D\}\ne u\{E(\theta\mid D)\}.

For example, a nonlinear penalty for a high adverse-event risk can make the posterior average differ materially from the utility at the average risk. The posterior distribution, not only its mean, then matters to the decision summary.

Worked Example: Expected Utility Across Posterior Scenarios

Suppose a hypothetical analysis uses the fixed utility function $u_t=10e_t-20h_t$, where $e_t$ is the probability of response and $h_t$ is the probability of a serious adverse event. Utility is measured in utility points. Three paired scenarios from the posterior have probabilities 0.20, 0.50, and 0.30. The paired values and utilities are:

Posterior scenario	Probability	$e_A$	$h_A$	$u_A=10e_A-20h_A$	$e_B$	$h_B$	$u_B=10e_B-20h_B$
1	0.20	0.90	0.05	8	0.60	0.15	3
2	0.50	0.70	0.15	4	0.80	0.15	5
3	0.30	0.50	0.35	-2	0.70	0.30	1

For example, in scenario 1, $u_A=10(0.90)-20(0.05)=9-1=8$ utility points. Average each treatment's utility over the posterior scenario probabilities:

$$ EU_A=0.20(8)+0.50(4)+0.30(-2)=1.60+2.00-0.60=3.00, $$

$$ EU_B=0.20(3)+0.50(5)+0.30(1)=0.60+2.50+0.30=3.40. $$

The posterior mean incremental utility for $A$ versus $B$ is $3.00-3.40=-0.40$ utility points. In scenario 1, $A$ is preferred because $8>3$; in scenarios 2 and 3, $B$ is preferred. Thus $P(u_A>u_B\mid D)=0.20$ and $P(u_B>u_A\mid D)=0.80$. The negative mean difference and the 0.20 probability of $A$ being preferred agree, but the two summaries describe different features of the posterior distribution.

Computing Posterior Expected Utility

For simple linear utility and available posterior means, expected utility may be calculated directly. Consider a hypothetical response-and-harm model with $u_t=100p_{R,t}-40p_{H,t}$, where $p_R$ and $p_H$ are response and harm probabilities. The coefficients are fixed utility weights in points per unit probability. Suppose independent binomial analyses with uniform Beta(1,1) priors give the following posterior distributions: for treatment $A$, $p_{R,A}\sim\mathrm{Beta}(18,22)$ and $p_{H,A}\sim\mathrm{Beta}(3,47)$; for treatment $B$, $p_{R,B}\sim\mathrm{Beta}(16,24)$ and $p_{H,B}\sim\mathrm{Beta}(2,48)$. These are illustrative posterior distributions, not clinical-study results.

For a $\mathrm{Beta}(a,b)$ distribution, the posterior mean is $a/(a+b)$. Because this utility function is linear, its posterior expectation is the same linear combination of posterior means:

Worked Example: Calculating Expected Utility From Beta Posteriors

For treatment $A$, the posterior mean response probability is $18/(18+22)=0.4500$, and the posterior mean harm probability is $3/(3+47)=0.06000$. Therefore,

EU_A=100(0.4500)-40(0.06000)=45.00-2.400=42.60\text{ utility points}.

For treatment $B$, the corresponding posterior means are $16/40=0.4000$ and $2/50=0.04000$. Thus,

EU_B=100(0.4000)-40(0.04000)=40.00-1.600=38.40\text{ utility points}.

The posterior mean incremental utility is $42.60-38.40=4.200$ utility points in favor of $A$. This calculation is valid for the stated linear utility even if response and harm parameters are dependent, because the expectation of a sum is the sum of expectations. It would not, by itself, provide the probability that $A$ is preferred; that probability depends on the posterior distribution of the difference, not only its mean.

For nonlinear utilities, or whenever the posterior distribution is available through simulation rather than a convenient formula, use posterior draws. For each draw, calculate each treatment's utility using the same specified function, then summarize those values. To compare treatments, calculate the difference within each draw. Pairing the treatment values from the same posterior draw preserves the dependence represented by that draw.

Obtain posterior draws.
Use draws that represent the fitted model's posterior distribution for the efficacy and safety parameters used by the utility model.

Evaluate utility draw by draw.
For each draw $m$, calculate $u_A^{(m)}$ and $u_B^{(m)}$ using the same outcome definitions, horizon, and fixed utility weights.

Average and compare.
Estimate each expected utility by the sample mean of its draw-specific utilities. Estimate the preference probability by the fraction of paired draws for which $u_A^{(m)}>u_B^{(m)}$.

Worked Example: Estimating a Treatment-Preference Probability

Suppose a hypothetical posterior simulation produces 10 paired draws of the incremental utility $\Delta^{(m)}=u_A^{(m)}-u_B^{(m)}$, in utility points. The values are $12, 8, 5, 3, 2, 1, -1, -4, -7,$ and $-9$. This small set is used to show the arithmetic, not to support a precise decision.

Six of the 10 differences are positive, so the estimated probability that $A$ has greater utility than $B$ is $6/10=0.600$. The mean incremental utility is

\overline{\Delta} =\frac{12+8+5+3+2+1-1-4-7-9}{10} =\frac{10}{10}=1.00\text{ utility point}.

Thus the simulated posterior mean favors $A$ by 1.00 utility point, while 60.0% of these posterior draws favor $A$. The probability estimate has substantial Monte Carlo error with only 10 draws. A rough simulation standard error is $\sqrt{0.600(1-0.600)/10}=0.1549$, which is large; a real analysis would use many more effective posterior draws and assess simulation stability.

Assumptions, Diagnostics, and Alternatives

A Bayesian expected-utility result is conditional on the model and decision specification. At minimum, the posterior must be an adequate representation of uncertainty in the parameters that feed the utility calculation; the utility function and weights must be specified on a meaningful, common scale; and the treatments must be compared for the same outcome definitions, time horizon, and target population. If utilities are built from outcome probabilities, those probabilities and the implied outcome distribution must be appropriate for the decision being considered. If posterior draws are used, their simulation must adequately explore the posterior. For a paired treatment comparison, use coherent paired draws from the joint posterior rather than independently pairing unrelated draws.

Useful diagnostics include checking posterior sampling and convergence, inspecting the range and shape of draw-specific utilities and differences, and verifying the utility calculation against hand calculations at selected draws. Check that probabilities remain within $[0,1]$, that event definitions are not inadvertently counted twice, and that the same weight convention is used for both treatments. Large changes in the estimated expected utility or preference probability across reasonable simulation runs can indicate inadequate Monte Carlo precision. A stable computation does not, however, validate a misspecified outcome model or an inappropriate utility scale.

A common alternative is a plug-in calculation: insert posterior mean parameters into the utility function and report the resulting value. This is equivalent to posterior expected utility for a linear function of the parameters, but not generally for a nonlinear function. The posterior mean also cannot determine a treatment-preference probability. Another alternative is to report only the posterior mean utility difference. That is useful for average utility but can conceal how often the difference changes sign. In contrast, the preference probability summarizes the posterior fraction of draws favoring one treatment under the fixed utility model; it does not quantify how large the advantage is on average.

The introductory material on stochastic multi-criteria acceptability analysis considers rankings across sampled preference weights. Here the focus is different: posterior uncertainty in efficacy and safety parameters is propagated through a specified utility function, with its preferences held fixed. If preference weights themselves are uncertain, that uncertainty is not included in the calculations above. Likewise, the results depend on the parameter model and on the selected attributes and utility weights. They should be presented as conditional decision summaries, not as model-free rankings or causal conclusions.

Common Mistakes

Using utility at the posterior mean as if it were always the posterior expected utility. This shortcut is generally justified only when the utility expression is linear in the uncertain parameters. For nonlinear utility, calculate utility over posterior draws or integrate over the posterior.
Confusing a mean with a probability of preference. A positive posterior mean difference does not imply that every draw favors the same treatment. Calculate the fraction of posterior draws with a positive difference to estimate the preference probability.
Interpreting the preference probability as the probability that a treatment is objectively best. It is conditional on the posterior model, the specified utility function, and the comparison being made.
Reporting a precise simulation estimate from too few effective draws. Increase the number of well-mixed posterior draws and examine Monte Carlo error; a large raw draw count does not fix poor sampling or model problems.

Key takeaway: Posterior expected utility averages a treatment's utility over posterior parameter uncertainty. The posterior probability of treatment preference is a separate summary, calculated from the proportion of posterior draws in which one treatment's utility exceeds the other's. Both are conditional on the model and the fixed utility specification.

Check Your Understanding

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

A treatment has posterior mean response probability 0.30 and a fixed linear response utility of 50 points per unit probability. What is its expected response utility?
For a nonlinear utility $u(p)=100p^2$, is $u\{E(p\mid D)\}$ generally equal to $E\{u(p)\mid D\}$? Explain.
Across 200 paired posterior draws, treatment $A$ has greater utility in 146 draws. Estimate the posterior probability that $A$ is preferred.
Why can a positive posterior mean incremental utility fail to show that one treatment is preferred in most posterior draws?

Solutions. (1) By linearity, expected utility is $50(0.30)=15.00$ utility points. (2) No. In general, $100\{E(p\mid D)\}^2\ne100E(p^2\mid D)$; the second expression includes posterior variation in $p$. (3) The estimate is $146/200=0.730$, or 73.0%. (4) A mean summarizes the magnitude and direction of differences, while the preference probability counts how often the difference is positive. A few large positive differences can produce a positive mean even if more than half of the draws are negative.