Risk Tolerance Thresholds for Clinical Trade-Offs

Why a Risk Tolerance Threshold Is Useful

A treatment may improve the chance of an efficacy outcome while also increasing the chance of a harmful outcome. A risk tolerance threshold makes that trade-off explicit: given a specified efficacy gain and a specified value assigned to each outcome, how large could the treatment harm rate be before its expected utility no longer exceeds control’s? This is a model-based comparison, not a universal clinical rule. Its answer depends on the outcomes, time horizon, utility values, and control risks used.

The introductory material on nonlinear utility for benefit and harm motivates representing outcomes by their value rather than treating every event as equally important. Here, consider a particular additive utility model and derive the maximum acceptable treatment harm rate from it. The resulting threshold is conditional on the model inputs; it does not by itself account for uncertainty in those inputs or establish a treatment recommendation.

Model and Assumptions

This formulation uses an additive utility: the utility contribution of efficacy is \(bE\), and the contribution of harm is \(-hH\). It assumes that these contributions add, that the utility values apply to both treatment groups, and that the specified efficacy and harm outcomes are defined over the same relevant population and time horizon. The utility scale can be rescaled without changing the threshold, provided both \(b\) and \(h\) are rescaled by the same positive factor. Their ratio, not their separate numerical scale, determines the trade-off.

The model also treats utility as depending on these outcome indicators alone. If other outcomes differ between arms or the value of one outcome depends on whether the other occurs, this two-term calculation may not represent the full comparison. Correlation between \(E\) and \(H\) does not alter the expectation of this particular additive utility: linearity of expectation gives \(E[U]=bP(E=1)-hP(H=1)\). However, dependence can matter for a different, non-additive utility model.

Deriving the Threshold

Assume first that the treatment efficacy gain \(\Delta p_E\) is specified and positive. The treatment is at least as good as control according to this model when its incremental expected utility is nonnegative. Solving that condition for the treatment harm probability gives the harm-rate threshold.

The right-hand side, \(p_{H,C}+(b/h)\Delta p_E\), is the uncapped break-even value. When it is at most 1, it is a feasible harm rate at which the model gives equal expected utility to treatment and control. Rates below it give positive incremental expected utility; rates above it give negative incremental expected utility, when those rates are feasible. The threshold includes the comparator’s harm rate because the decision depends on the increase in harm, not on the treatment harm rate in isolation.

A harm probability cannot exceed 1. If the uncapped break-even value is greater than 1, no feasible treatment harm rate reaches break-even. Even at \(p_{H,T}=1\), incremental expected utility remains positive under the stated model. In this case the maximum feasible treatment harm rate is 1, but 1 is not an indifference point. If the uncapped value equals 1, by contrast, a treatment harm rate of 1 is a feasible break-even point. These cases must not be conflated.

For a negative efficacy gain, the same inequality still applies algebraically, but it does not describe a positive efficacy benefit being traded against harm. The examples below use a positive specified gain. If efficacy and harm risks are uncertain rather than fixed, this calculation is a threshold conditional on chosen values; uncertainty requires separate analysis.

Why the Derivation Works

The threshold follows directly from comparing expected utilities under the same utility function. For each arm, the expected utility is the efficacy value multiplied by its event probability, minus the harm loss multiplied by its event probability. Subtracting control from treatment cancels any terms shared by both arms and leaves only the differences in efficacy and harm risks:

Setting this difference equal to zero identifies the break-even condition. Rearranging the inequality for nonnegative incremental utility gives \(p_{H,T}\leq p_{H,C}+(b/h)\Delta p_E\). This also shows the direction of the trade-off: for fixed positive weights, a larger efficacy gain increases the harm rate compatible with break-even; a larger harm loss relative to efficacy value reduces the tolerated increase.

Diagnostics, Alternatives, and Limitations

Before using a threshold, check that efficacy and harm are defined consistently across arms and over the same period. A difference between a 12-month treatment efficacy rate and a lifetime control harm rate does not represent the stated model comparison. Check that each probability is an absolute risk, that the efficacy gain is treatment minus control, and that the threshold is not confused with the allowed increase in harm. A direct substitution into \(\Delta EU=b\Delta p_E-h(p_{H,T}-p_{H,C})\) is a useful arithmetic and sign check.

The utility values deserve particular scrutiny. They express how much the decision-maker values the efficacy outcome relative to the harm, not how frequently either event occurs. If the utility ratio \(b/h\) is poorly supported, the resulting threshold may look precise while being highly preference-dependent. A threshold calculated from estimated event rates and elicited utilities should therefore be presented with its assumptions and input values; a single plug-in value does not express uncertainty in those inputs.

A risk-difference criterion is the closest arithmetic alternative: it states a tolerated increase in harm risk directly. In this additive model that increase at break-even is \((b/h)\Delta p_E\), so the utility model explains how the permitted difference follows from the efficacy gain and relative outcome values. A risk-ratio criterion instead compares \(p_{H,T}/p_{H,C}\). It can be useful when proportional changes are the chosen decision scale, but it is not equivalent to this utility-derived threshold: the same risk ratio implies different absolute harm increases at different control risks. Neither a fixed risk difference nor a fixed risk ratio automatically represents the specified utility trade-off.

The result is limited to the utility structure and inputs used. It is not a causal estimate of treatment effects, a guarantee about an individual patient, or a statement that harms below the threshold are harmless. If utility is nonlinear in event probability, if an event’s value depends on another outcome, or if relevant outcomes differ beyond the two represented here, the two-term equation may be inadequate. The analysis also does not make uncertain inputs certain: the threshold answers a conditional question for the specified values, while uncertainty and preference variation require explicit consideration.

Common Pitfalls

• Calling a value above 1 an indifference point. A value such as 1.04 cannot be a probability and is not a feasible tie. If the uncapped value exceeds 1, check utility at the feasible maximum \(p_{H,T}=1\); if incremental utility is still positive, all feasible rates remain above break-even in utility.
• Reporting only the treatment rate or only the risk increase. The threshold equation gives a treatment harm rate. Subtracting \(p_{H,C}\) gives the corresponding increase over control. State which one is being reported.
• Treating a model threshold as a clinical standard. The threshold changes with the efficacy gain and the relative utility values. It is a conditional result, not a universally acceptable rate.
• Reversing the efficacy difference. Use \(p_{E,T}-p_{E,C}\). A reversed subtraction changes the sign of the benefit term and can reverse the conclusion.
• Ignoring the outcome period or population. Probabilities from different follow-up windows or populations cannot be combined as if they described the same comparison without additional justification.