From Outcome Contrasts to Net Clinical Benefit
A treatment can reduce the risk of stroke while increasing the risk of bleeding. Reporting those outcomes separately is important, but it does not by itself express their overall balance according to a stated set of preferences. Net clinical benefit summarizes the weighted balance of specified benefits and harms relative to a reference treatment. Its value depends on the outcome contrasts, the preference weights, the population, and the time horizon; it is not an objective property of a drug independent of those choices.
The introductory material on utility functions and patient preferences provides the basis for assigning values to clinical outcomes. Here the task is narrower: combine estimated differences in efficacy and safety on one common preference-weighted scale. The result is meaningful only when the outcomes, directions, units, and weights are clearly specified.
For stroke prevention and bleeding risk, let \(N_S\) and \(N_B\) denote the number of stroke and major-bleeding events, respectively, over the stated horizon. Define the expected stroke reduction and expected bleeding increase per patient as
where \(T\) is the treatment and \(R\) is the reference. Thus, \(d_S>0\) means fewer strokes with treatment, while \(d_B>0\) means more major bleeds with treatment. If \(w_S\) is the preference value of preventing one stroke and \(w_B\) is the preference cost of one additional major bleed, then
The weights must share a scale, such as quality-adjusted life-years (QALYs) per event or preference points per event. If contrasts are risks per patient and weights are QALYs per event, NCB is QALYs per patient over the horizon. If contrasts are expected event counts per 1,000 patients, NCB is weighted QALYs per 1,000 patients.
Assumptions and a Short Derivation
This weighted sum is appropriate when the analysis has a well-defined population and time horizon, and when each outcome contrast refers to that same population and horizon. The direction and unit of every contrast must be explicit. The weights must represent the relative values of the outcomes on a common scale. In addition, the additive calculation assumes that the combined value of outcomes can be represented by adding their weighted contributions. If the value of a stroke depends on whether a patient also experiences a bleed, a simple additive score may not represent that interaction.
To see why the formula has this form, suppose each stroke contributes a utility loss of \(w_S\) and each major bleed contributes a utility loss of \(w_B\). Under an additive outcome-value model, expected utility for treatment \(T\) minus expected utility for reference \(R\) is
The last equality follows from the definitions: \(E(N_{B,R})-E(N_{B,T})=-d_B\). This derivation also shows why sign discipline matters. An increase in a harmful outcome contributes negatively, while a reduction in that outcome contributes positively. If the outcomes are binary over the horizon, their event probabilities can be used as expected event counts per patient. For recurrent events, expected counts rather than the probability of at least one event may be needed.
Worked Example: Stroke Reduction With an Increase in Major Bleeding
In a hypothetical two-year comparison, treatment is associated with 18 fewer strokes and 6 additional major bleeds per 1,000 patients than the reference. Suppose the preference weights are 0.30 QALY per stroke prevented and 0.08 QALY lost per additional major bleed. Assume the contrasts are on the same population and horizon and the additive weighted model is appropriate.
The contrasts are \(d_S=18\) events per 1,000 patients and \(d_B=6\) events per 1,000 patients. The weighted calculation is
Under these specified weights, the balance is positive: the stroke-prevention contribution exceeds the weighted bleeding harm by 4.92 QALYs per 1,000 patients over two years. Dividing by 1,000 gives 0.00492 QALY per patient over the same horizon. This is a weighted summary, not a claim that every patient gains that amount or that the treatment is preferable under all possible preferences.
Reading the Result and Checking the Trade-Off
A positive NCB means that the weighted contributions favor treatment over the reference under the chosen model and weights; a negative value means they favor the reference. Zero is the break-even point for those particular weights and contrasts. The size of the number has meaning only with its scale and unit attached. A score of 4.92 QALYs per 1,000 patients is not directly comparable with 4.92 preference points per patient.
The calculation can also show how a conclusion depends on preferences. Setting NCB to zero and solving for the bleeding weight gives a break-even value, provided \(d_B>0\):
At this value, the weighted stroke benefit and bleeding harm exactly balance. This is a threshold in the preference scale, not a probability and not a statistical significance threshold.
Worked Example: Finding the Break-Even Bleeding Weight
In a hypothetical one-year analysis, treatment prevents 12 strokes and causes 10 additional major bleeds per 1,000 patients. Let the stroke-prevention weight be 0.25 QALY per event and the bleeding weight be 0.10 QALY per event. First calculate NCB at the stated bleeding weight, then find the break-even bleeding weight.
At the stated weights,
For the break-even value, set \(w_Sd_S-w_Bd_B=0\) and solve:
With all else held fixed, a bleeding weight below 0.300 QALY per event gives positive NCB, while a weight above 0.300 gives negative NCB. At the stated weight of 0.10, NCB is positive. This threshold makes the preference trade-off transparent; it does not establish which bleeding weight is appropriate for a particular patient or population.
Worked Example: Both Stroke and Bleeding Risks Decrease
In a hypothetical comparison of 1,000 patients over 18 months, the reference group has 40 strokes and 18 major bleeds, while the treatment group has 25 strokes and 12 major bleeds. Let the weights be 0.20 QALY per stroke and 0.05 QALY per major bleed.
The stroke reduction is \(40-25=15\) events per 1,000. Because bleeding also decreases, the bleeding contrast defined as treatment minus reference is \(12-18=-6\) events per 1,000. Therefore,
The negative bleeding contrast correctly produces a positive contribution: six fewer bleeds add 0.30 weighted QALY per 1,000 patients under the stated weights. Reversing the sign convention without changing the formula would incorrectly turn this safety improvement into a harm.
Diagnostics, Alternatives, and Limitations
Before interpreting NCB, check the calculation and its inputs. Confirm that treatment and reference are named consistently; that each contrast uses the declared direction; and that event units, population, and follow-up horizon match. Inspect whether risk differences are reported per patient, per 100, or per 1,000, and apply the same scale to both outcomes. Check the sign with a simple scenario: if treatment reduces both stroke and bleeding, both contributions should favor treatment. If treatment increases both, both should count against it.
The estimated contrasts also need an appropriate basis. A descriptive difference between groups does not automatically represent a treatment effect: its interpretation depends on the study design and the assumptions supporting the comparison. Likewise, a positive point estimate of NCB does not by itself communicate the precision of the estimates or uncertainty in the preferences. Sensitivity analysis over plausible weights can show whether the sign or practical size of NCB is robust, but a single weighted score can conceal the separate outcome changes.
A weighted sum is not the only way to present benefit and harm. Reporting each outcome contrast separately preserves their clinical meaning and avoids imposing a single preference trade-off. A composite endpoint combines events into a defined endpoint, but its result depends on which events are included and how the composite is analyzed; it does not necessarily assign explicit preference values to different components. NCB instead makes the chosen trade-off visible through the weights. Neither presentation replaces the other: reporting the component outcomes alongside NCB helps readers see what drives the summary.
Important limitations include the dependence of results on elicited or selected weights, possible variation in preferences across patients, and the simplifying assumption that outcome values add without interaction. A score can also be misleading if outcomes overlap or if event counts and risks are mixed without a clear estimand. The weighted sum of expected outcomes is valid under the stated additive model even when outcomes are correlated, but correlation and overlap can matter for uncertainty calculations and for whether the chosen outcome definitions represent distinct consequences. NCB should therefore be reported with its reference, population, horizon, outcome definitions, weights, contrast units, and component contrasts.
Common Pitfall
Name the treatment, reference, population, outcomes, and time horizon.
Define reductions in harmful outcomes as positive benefits, or define all contrasts in a consistent treatment-minus-reference direction and account for outcome harms in the formula.
Use preference weights on a common scale and retain the units through the arithmetic.
State the weighted result with its unit, show component contributions, and describe how the chosen weights affect the conclusion.
Check Your Understanding
Answer each question using the stated sign convention and units.
- A treatment prevents 8 strokes and causes 5 additional major bleeds per 1,000 patients. The weights are 0.20 QALY per stroke and 0.06 QALY per bleed. What is NCB per 1,000 patients?
- In a comparison, both stroke risk and major-bleeding risk are lower with treatment. What sign should the bleeding contrast have under the definition \(d_B=E(N_{B,T})-E(N_{B,R})\), and how does it enter the NCB formula?
- A weighted analysis reports NCB of \(-1.2\) QALYs per 1,000 patients. What does the sign mean, and what does it not establish?
- Why is it invalid to subtract the number of strokes prevented from the number of additional bleeds without weights?
Solutions. (1) \(\mathrm{NCB}=(0.20)(8)-(0.06)(5)=1.60-0.30=1.30\) QALYs per 1,000 patients over the specified horizon. (2) The bleeding contrast is negative because treatment has fewer bleeds, so \(-w_Bd_B\) is positive and represents a weighted benefit. (3) Under the stated weights and model, the weighted balance favors the reference by 1.2 QALYs per 1,000 patients. The sign alone does not establish statistical certainty, universal preference, or that every patient is worse off. (4) Stroke and bleed events are not equivalent units; subtraction without weights implicitly assigns them equal value. A common preference scale is needed to combine them meaningfully.