Setting Up an Oncology Benefit-Risk Comparison
A treatment that produces more tumor responses may also cause more severe toxicity. Survival can favor one regimen while response favors another, and the balance depends partly on how patients value these outcomes. Expected utility makes that trade-off explicit: it combines outcome probabilities with stated preferences to produce a summary score for each regimen.
The calculations below use hypothetical data for two oncology regimens, A and B, in the same target population. Outcomes are assessed over a common 12-month period. Tumor response is classified as complete response (CR), partial response (PR), stable disease (SD), or progressive disease (PD). Survival is represented by whether a patient is alive at 12 months. Toxicity is represented by the highest grade experienced during the period, on a scale from 0 to 4. The response categories are mutually exclusive, as are the toxicity grades.
The model used here is additive. It assigns a score to tumor response, adds a weighted contribution for 12-month survival, and adds a grade-specific toxicity score. These are preference scores, not measured clinical quantities. For example, the response score of 1 for CR does not mean that a complete response is literally one unit of health. It is a value on the chosen scale.
In this expression, \(r\) is the response score, \(s\) equals 1 if the patient is alive at 12 months and 0 otherwise, and \(q_g\) is the utility contribution assigned to the patient's highest toxicity grade \(g\). The coefficients \(w_R\) and \(w_S\) are preference weights; \(\lambda\) scales the toxicity contribution. The example uses these response scores and toxicity contributions:
| Outcome category | Score or contribution |
|---|---|
| CR | 1.00 |
| PR | 0.80 |
| SD | 0.30 |
| PD | 0.00 |
| Toxicity grade 0 | 0.00 |
| Toxicity grade 1 | -0.05 |
| Toxicity grade 2 | -0.15 |
| Toxicity grade 3 | -0.40 |
| Toxicity grade 4 | -0.80 |
For the base case, \(w_R=1.00\), \(w_S=0.50\), and \(\lambda=1.00\). These values represent one hypothetical set of preferences. They are not universal weights for oncology patients.
Worked Example: Calculating the Outcome Scores
Worked Example: Calculating the Outcome Scores
The following hypothetical probabilities describe the two regimens. Each response distribution and each toxicity distribution sums to 1.00. Survival is the probability of being alive at 12 months.
| Regimen | CR | PR | SD | PD | Alive at 12 months |
|---|---|---|---|---|---|
| A | 0.12 | 0.38 | 0.30 | 0.20 | 0.68 |
| B | 0.08 | 0.32 | 0.36 | 0.24 | 0.62 |
| Regimen | Grade 0 | Grade 1 | Grade 2 | Grade 3 | Grade 4 |
|---|---|---|---|---|---|
| A | 0.30 | 0.30 | 0.22 | 0.13 | 0.05 |
| B | 0.45 | 0.28 | 0.17 | 0.08 | 0.02 |
Response score for A. Multiply each category probability by its score and add:
Response score for B.
Expected toxicity contribution. The grade-specific contributions are negative for grades 1 through 4. For A, the expected contribution is \((0.30)(0)+(0.30)(-0.05)+(0.22)(-0.15)+(0.13)(-0.40)+(0.05)(-0.80)=-0.140\). For B, it is \((0.45)(0)+(0.28)(-0.05)+(0.17)(-0.15)+(0.08)(-0.40)+(0.02)(-0.80)=-0.0875\). Thus B has the less negative expected toxicity contribution under these assigned scores.
Combining the Attributes
Expected utility can now be calculated using the base-case weights. The survival contribution is the survival probability multiplied by \(w_S=0.50\). The toxicity contribution uses \(\lambda=1.00\), so the expected toxicity scores calculated above enter without further scaling.
Worked Example: Comparing Base-Case Expected Utilities
For regimen A, the response contribution is \((1.00)(0.514)=0.514\), the survival contribution is \((0.50)(0.68)=0.340\), and the expected toxicity contribution is \(-0.140\). Therefore:
For regimen B, the response contribution is \((1.00)(0.444)=0.444\), the survival contribution is \((0.50)(0.62)=0.310\), and the expected toxicity contribution is \(-0.0875\). Therefore:
Under these hypothetical probabilities and preferences, regimen A has the higher expected utility: \(0.714-0.6665=0.0475\) utility points. The difference reflects the model's specified trade-offs: A has higher expected response and 12-month survival, while B has a less negative toxicity contribution. This conclusion is conditional on the chosen outcome definitions, probabilities, and weights.
Why the Additive Calculation Works
The expected utility does not require response, survival, and toxicity to be statistically independent. The key assumption is that the specified utility for a patient's combined outcome is additive in these attributes. To see the implication, let \(X\) denote the complete joint outcome, containing \(r(X)\), \(s(X)\), and \(q_g(X)\). By the definition of expectation and the additive utility function:
The last equality follows from linearity of expectation. It holds whether the attributes are independent, positively associated, or negatively associated. Since \(s\) is a binary indicator, \(E[s\mid a]=P(s=1\mid a)\). Since response categories and toxicity grades are exhaustive and mutually exclusive within their respective attributes, their expected scores are the sums of category probabilities multiplied by the corresponding scores.
This algebraic result does not establish that an additive utility function is appropriate for every patient or setting. The model must represent the intended preferences. In particular, response and survival may describe related aspects of disease control. Their contributions should be combined only if the stated utility model is intended to value both attributes in this way. A calculation cannot resolve an inappropriate or poorly elicited preference structure.
Worked Example: When the Preference Weight Changes
Worked Example: When the Preference Weight Changes
Suppose the response and survival weights remain \(w_R=1.00\) and \(w_S=0.50\), but the toxicity scale \(\lambda\) is allowed to vary. Regimen A's response-plus-survival contribution is \(0.514+0.340=0.854\); B's is \(0.444+0.310=0.754\). The expected toxicity contributions before scaling are \(-0.140\) for A and \(-0.0875\) for B. Thus:
The regimens tie when the difference equals zero. Solving gives:
For \(\lambda=1.00\), the difference is \(0.100-0.0525=0.0475\), so A is preferred, as in the base case. For \(\lambda=2.00\), the difference is \(0.100-0.105=-0.005\), so B is preferred. With these other inputs fixed, a toxicity multiplier below 1.905 favors A, a multiplier above 1.905 favors B, and the calculated utilities are equal at 1.905. This threshold describes the model's sensitivity to a preference weight; it is not a biological or clinical threshold.
Assumptions, Diagnostics, and Alternatives
Before interpreting a result, check the inputs and the structure of the comparison. The calculations require that the regimens refer to the same target population and the same assessment period; category definitions are consistent; and probabilities are on a common basis. Each distribution must be complete and sum to 1, up to rounding. The response and toxicity scores and the weights must also be specified on a scale where their sum has the intended interpretation.
Confirm the response categories, survival time point, toxicity grading, and toxicity summary rule are clear and applied consistently to both regimens.
Verify that each set of category probabilities is nonnegative and sums to 1. Investigate missing categories or discrepancies larger than expected rounding error.
Ask whether the category scores and weights reflect the intended decision perspective and whether the additive form represents the intended trade-offs.
Recalculate across plausible preference weights and outcome probabilities. If modest changes reverse the ranking, report that the choice is sensitive rather than presenting one regimen as an unconditional winner.
Common warning signs include toxicity probabilities that do not sum to 1, different follow-up windows across attributes, unclear handling of patients with missing outcomes, and a preferred regimen that changes under small, defensible changes in weights. A further limitation is that a point-estimate calculation, such as the one above, does not show uncertainty in the clinical probabilities or preferences. When those inputs are uncertain, results should be examined over plausible scenarios or distributions rather than treated as exact.
An additive score is transparent and easy to decompose, but it assumes the stated additive trade-offs. A multiplicative utility model can represent interactions between attributes, but requires a different utility structure and additional preference information. Reporting outcomes separately is a useful alternative when no defensible common preference scale is available; it preserves the trade-offs for decision makers but does not produce a single ranking. The appropriate approach depends on the decision question and the quality of the evidence and preference inputs.
Common Mistake
Check Your Understanding
Use the hypothetical data and base-case model above. Solutions are included after the questions.
- If a new hypothetical regimen has response probabilities CR 0.10, PR 0.40, SD 0.30, and PD 0.20, what is its expected response score?
- Why can the additive expected-utility calculation use marginal response, survival, and toxicity distributions without assuming those attributes are independent?
- At \(\lambda=2.00\), which regimen has the higher expected utility, and by how much?
- Name two checks that should be completed before interpreting a calculated regimen ranking.
Solutions. (1) The expected response score is \((0.10)(1.00)+(0.40)(0.80)+(0.30)(0.30)+(0.20)(0.00)=0.100+0.320+0.090=0.510\). (2) Under the additive utility function, linearity of expectation gives the expected total as the sum of each weighted attribute's expectation, regardless of dependence. (3) The utility difference is \(EU(A)-EU(B)=0.100-(0.0525)(2.00)=-0.005\), so B is higher by 0.005 utility points. (4) For example, verify that category probabilities are nonnegative and sum to 1, and confirm that both regimens use the same population, outcome definitions, and assessment period. The response, survival, and toxicity scores and weights should also be checked for consistency with the intended preference model.