Correlated Outcomes in Benefit-Risk Assessment

Why the Joint Distribution Matters

A patient may experience both a treatment benefit and a toxicity during the same follow-up period. Summarizing the probability of benefit and the probability of toxicity separately does not tell us how often these outcomes occur together. That overlap matters when the value assigned to a benefit depends on whether toxicity also occurred, or when utility is otherwise non-additive.

The earlier material on averaging utilities establishes the basic calculation: multiply each possible outcome’s utility by its probability and sum. For efficacy and toxicity assessed on the same patients over the same horizon, the relevant probabilities may be joint probabilities. Replacing them with probabilities constructed as if the outcomes were independent can change expected utility—and, in some cases, the preferred treatment.

For discrete outcomes, expected utility is calculated over the joint distribution:

This calculation assumes the outcome categories and their utilities are defined on a common, clinically meaningful assessment horizon; the joint probabilities refer to the same target population; and the probabilities form a valid distribution (they are nonnegative and sum to one). The utility function must also represent the stated preferences for each possible outcome pair. If outcomes are continuous, the corresponding calculation integrates utility over their joint density.

Dependence and Utility Interaction

Consider binary indicators: \(B=1\) means the patient experiences the specified efficacy outcome, and \(T=1\) means the patient experiences the specified toxicity. Let \(p_B=P(B=1)\), \(p_T=P(T=1)\), and \(q=P(B=1,T=1)\). The four joint probabilities are \(q\), \(p_B-q\), \(p_T-q\), and \(1-p_B-p_T+q\). Their nonnegativity implies:

Thus, knowing both marginal probabilities generally does not determine how often benefit and toxicity occur together. Independence supplies one particular value, \(q=p_Bp_T\), but it is an assumption, not a consequence of knowing the two marginals.

For binary outcomes, any utility table can be written as \(U(B,T)=a+bB+cT+dBT\), where \(a\), \(b\), \(c\), and \(d\) are determined by its four utility values. Specifically, \(a=U(0,0)\), \(b=U(1,0)-U(0,0)\), \(c=U(0,1)-U(0,0)\), and \(d=U(1,1)-U(1,0)-U(0,1)+U(0,0)\). Taking expectations gives:

The coefficient \(d\) measures the interaction in the utility table: it is the extra change in utility when both outcomes occur, beyond the changes represented by the separate benefit and toxicity terms. Since \(\operatorname{Cov}(B,T)=q-p_Bp_T\), the difference between expected utility under the actual joint distribution and under independence is:

For binary outcomes with nonzero variances, \(\operatorname{Cov}(B,T)=\rho\sqrt{p_B(1-p_B)p_T(1-p_T)}\), where \(\rho\) is their correlation. This identity shows why correlation alone is not enough to determine the direction or size of the utility change: both the utility interaction and the dependence matter.

When Dependence Does Not Change Expected Utility

It is important not to conclude that correlation always changes expected utility. If utility is additive in benefit and toxicity, then \(d=0\). In that case, expected utility depends only on the marginal probabilities, even when the outcomes are correlated. This is a mathematical property of additive utility—not evidence that the outcomes are independent.

Checking the Direction and Consequences of Dependence

When utility includes an interaction, positive dependence can raise or lower expected utility depending on the sign of \(d\). In the first example, the combined outcome had a lower utility than would be implied by adding the separate benefit and toxicity contributions, so \(d\) was negative. Positive covariance then reduced expected utility relative to independence. With a positive interaction, the same positive covariance would increase expected utility.

Modeling and Diagnosing the Joint Distribution

A joint distribution may be estimated directly from paired outcomes, or represented through a marginal model and a conditional model. For example, the identity \(P(B,T)=P(B)P(T\mid B)\) is valid by the definition of conditional probability when \(P(B)>0\). Setting \(P(T\mid B)=P(T)\) for all relevant values of \(B\) imposes independence. The chosen representation should match the available data and make the dependence assumption explicit.

Useful checks include confirming that cell probabilities are nonnegative and sum to one, verifying that row and column totals reproduce the intended marginals, and comparing the estimated co-occurrence with the independence prediction. For binary outcomes, compare \(q\) with \(p_Bp_T\); for more outcome categories, inspect the joint table or an appropriate dependence summary. Sparse or empty cells can make joint estimates unstable, especially for uncommon outcomes. A fitted joint model should therefore be checked for implausible probabilities and sensitivity to reasonable alternative dependence assumptions.

If only the marginal probabilities are available, the dependence is not identified from those marginals alone. The feasible range for \(q\) can still be used to assess how much expected utility could vary under different joint distributions. In the examples, \(p_B=0.60\) and \(p_T=0.30\) allow \(0\leq q\leq0.30\). For the interacting utility \(10B-8T-6BT\), expected utility is \(3.60-6q\), so this range permits values from 3.60 points at \(q=0\) to 1.80 points at \(q=0.30\). This is a dependence sensitivity range, not a confidence interval.

A joint model is not automatically preferable in every setting. It can require more data and more modeling assumptions than separate marginal models, while sparse joint cells may be difficult to estimate reliably. Conversely, assuming independence for convenience can conceal clinically important co-occurrence. Association between outcomes also does not establish that one outcome causes the other; this calculation describes their joint distribution and its implications for utility, not a causal mechanism.