Binary Endpoint Utility and Number Needed to Treat

From a Binary Response to a Dose Decision

A binary endpoint records whether a defined event occurred by a specified time: for example, whether a patient achieved a prespecified response by week 12. For dose selection, a response probability alone may not be enough. The decision also depends on the value assigned to response versus nonresponse and on how much the probability changes relative to a meaningful comparator. Expected utility summarizes the first question; the number needed to treat (NNT) expresses an absolute response difference in a familiar frequency scale.

These quantities answer different questions. Expected utility is an average value on a chosen utility scale. NNT is the reciprocal of an absolute difference in response probability, expressed as the number treated at one dose rather than a comparator to obtain one additional response on average. Neither should be interpreted without specifying the endpoint, time horizon, population, and comparator.

The expression assumes that the binary endpoint adequately represents the outcome being valued and that the utilities \(u_1\) and \(u_0\) apply to the population and decision in question. The probability must refer to a defined estimand: the same response criterion, assessment time, and target population at every dose. To interpret the contrast causally, the dose-specific probabilities must represent outcomes under the corresponding dose interventions. A randomized comparison can support this interpretation under appropriate conduct and outcome ascertainment; nonrandomized data require assumptions sufficient to identify those intervention probabilities.

This is a deliberately simple utility specification. It assigns one utility to response and another to nonresponse, without distinguishing different kinds or severities of outcomes within either state. If important benefits or harms are omitted, the resulting utility is not a complete summary of benefit and risk.

Deriving Expected Utility and NNT

For a binary outcome, expected utility follows directly by weighting each possible utility by its probability. The incremental utility has a particularly simple form:

Thus, under this two-state utility, the change in expected utility is the absolute response-probability difference multiplied by the utility gap between response and nonresponse. If response is valued more than nonresponse, \(u_1>u_0\), a dose with higher response probability also has higher expected utility. The size of the utility difference depends on the chosen scale; the response-probability difference does not.

For a beneficial response, define the absolute response difference against the reference as \(RD=p(d)-p(d_0)\). When \(RD>0\), the NNT is the reciprocal of that difference. The reason is that in a group of \(m\) patients treated at each strategy, the expected difference in the number responding is \(m p(d)-m p(d_0)=mRD\). Setting this expected excess equal to one gives \(m=1/RD\).

NNT is usually reported as a positive number. If a whole number is required for a practical statement, round upward: a fractional value such as 6.67 means that, on average, treating about seven patients at the candidate dose rather than the reference dose yields one additional response. This is a group-average interpretation, not a claim that a particular seventh patient will respond because of treatment.

The NNT is not the number needed to produce one response in total. It refers to one additional response compared with the stated reference. It also is not an individual treatment effect: the difference compares average response probabilities under two strategies.

Comparing Candidate Doses

When selecting among doses, calculate each candidate's response probability relative to the same reference, then show the response difference and the corresponding expected-utility difference. This prevents a common error: comparing NNT values without checking their underlying probabilities and comparators. A smaller NNT for a beneficial response means a larger absolute response difference, but it does not by itself establish that the dose is preferable. Other valued outcomes, constraints, or dose-related consequences may also matter.

Direction, Uncertainty, and Diagnostics

The sign of the absolute difference must be checked before calculating an NNT for benefit. If the endpoint is a desirable response and \(p(d)<p(d_0)\), the candidate dose has fewer responses; \(1/\{p(d)-p(d_0)\}\) is negative and should not be reported as a beneficial NNT. If the endpoint instead records an undesirable event, define the absolute reduction as \(p(d_0)-p(d)\). A positive reduction can be expressed as the number needed to treat to prevent one event, provided the event definition and comparison are explicit.

NNT also depends strongly on the absolute difference. A difference close to zero produces a very large NNT; at zero, the reciprocal is undefined. When estimates are uncertain, uncertainty in the difference must be carried through the reciprocal transformation. If an interval for the absolute difference includes zero, the corresponding NNT uncertainty crosses the point where the reciprocal is undefined; a single finite interval can therefore be misleading. Report the probability difference and its uncertainty alongside any NNT, rather than presenting NNT alone.

In a dose-finding analysis, the probabilities may be model-based predictions rather than observed proportions. Check that the predicted probabilities are within the valid range from 0 to 1, refer to the same estimand, and behave plausibly across the candidate doses. Inspect the fit and uncertainty of the dose-response model: a precise-looking NNT calculated from poorly supported extrapolation is not reliable. Also verify that the reference dose is clinically interpretable and that changes in endpoint definition or follow-up have not been mixed across dose comparisons.

How NNT Complements Other Summaries

Response probability, risk difference, relative risk, and expected utility provide related but noninterchangeable summaries. The probability at a dose gives the modeled response level. The risk difference gives the absolute change against a comparator and is the quantity from which NNT is calculated. A relative risk gives a ratio of probabilities; it can describe proportional change but does not determine the absolute gain without the baseline probability. Expected utility incorporates the value assigned to the outcome states. In the two-state case, the incremental utility equals the risk difference multiplied by the utility gap, so NNT and utility are directly connected only when that gap is specified.

For example, the same relative increase in response can correspond to very different absolute differences at different baseline probabilities, and therefore different NNTs. NNT is also not invariant to the chosen comparator, response threshold, or observation period. The standard name “number needed to treat” should not obscure these dependencies: always state “versus what,” “for which outcome,” and “over what time.”

Common Pitfalls and Limits

• Using a relative difference in the NNT formula. The denominator is the absolute probability difference, not a risk ratio or percentage increase relative to baseline.
• Omitting the comparator. An NNT without a reference strategy is incomplete because changing the comparator changes the absolute difference.
• Treating NNT as a patient-level prediction. It summarizes an average contrast across groups; it does not identify which individuals benefit.
• Ignoring direction. A negative response difference indicates fewer beneficial responses at the candidate dose, not a negative beneficial NNT.
• Reporting a huge NNT without uncertainty. When the difference is small or imprecise, its reciprocal can be extremely unstable. Show the probability difference and its uncertainty.
• Assuming a single binary endpoint captures all value. The calculation is incomplete if relevant benefits or harms are left out of the endpoint and utility specification.

For dose decisions, use NNT as a translation of an absolute response contrast, not as a stand-alone ranking rule. State the dose and comparator, the endpoint and time horizon, the modeled probabilities, and the utility assumptions. If uncertainty or omitted outcomes could change the decision, make those limitations visible rather than allowing a simple reciprocal to imply certainty.