Number needed to treat, explained: what one number hides
Number needed to treat explained: how NNT comes from absolute risk, why it depends on baseline risk and time frame, and pitfalls researchers warn about.
By Manouchehr Hessabi, MD, MPH
The number needed to treat (NNT) is the expected number of people who would need to receive one intervention instead of a comparator for one additional person to avoid, or to experience, an outcome over a stated period of time. It is the inverse of the absolute risk reduction, which is why it turns an abstract percentage back into a count of people (Cook and Sackett, 1995). Its usefulness depends entirely on three things that the single number does not show on its own: the comparator, the time frame and the baseline risk of the people studied.
What is the number needed to treat?
The Cochrane Handbook for Systematic Reviews of Interventions, the standard methods reference for systematic reviews, defines the NNT as the expected number of people who must receive the experimental rather than the comparator intervention for one additional person to either incur or avoid an event "in a given time frame" (Cochrane Handbook, chapter 15).
The measure was promoted as a way to make trial results easier to weigh. A 1995 article in the BMJ by Richard Cook and David Sackett described it as more meaningful for clinical decisions than the relative risk, the relative risk reduction or the odds ratio, and noted that it "conveys both statistical and clinical significance" (Cook and Sackett, 1995). That article cites an earlier paper by Laupacis, Sackett and Roberts in the New England Journal of Medicine, an assessment of clinically useful measures of the consequences of treatment (Laupacis, Sackett and Roberts, 1988).
A few terms help before going further. Risk here means the proportion of a group that experiences an event over a period. The absolute risk reduction, also called the risk difference, is the plain difference between the risk in one group and the risk in the other. The relative risk reduction expresses that difference as a share of the starting risk. The explainer on relative versus absolute risk covers why those two framings can leave such different impressions.
How is the NNT calculated from a trial result?
The arithmetic is short. Cook and Sackett describe the NNT as calculated on the inverse of the absolute risk reduction, so the NNT is 1 divided by the risk difference.
Take a purely hypothetical example, with round numbers chosen for illustration and not drawn from any study. Suppose that over one year, 10 percent of people in a comparator group have an event and 8 percent of people in a treated group do.
- The absolute risk reduction is 10 percent minus 8 percent, which is 2 percentage points, or 0.02.
- The NNT is 1 divided by 0.02, which is 50.
- Read in full: about 50 people would need to receive the treatment rather than the comparator for one year for one additional person to avoid the event.
The same hypothetical result can also be described as a relative risk reduction of 20 percent (2 points out of a starting 10). Both descriptions are correct. The NNT simply keeps the absolute size of the effect in view.
Two conventions are worth knowing. Cochrane states that NNTs are always expressed as positive whole numbers, and that by convention computed values are rounded up to the next whole number (Cochrane Handbook, chapter 15). Rounding up is the conservative choice: it never makes a treatment look more efficient than the data suggest.
Why does the same treatment have different NNTs?
Because the NNT is built from an absolute difference, anything that changes the absolute difference changes the NNT. Three factors matter most.
Baseline risk. Cochrane illustrates this with oral anticoagulation in atrial fibrillation, drawing on a Cochrane review by Aguilar and Hart (2005). In higher-risk patients, with stroke rates of about 12 percent (120 per 1000) per year, warfarin prevents about 70 strokes per 1000 patients per year. In lower-risk patients, with a stroke rate of about 2 percent (20 per 1000) per year, it prevents only about 12 per 1000 (Cochrane Handbook, chapter 15).
Converting those rounded figures into NNTs shows the scale of the difference. About 70 events avoided per 1000 corresponds to an NNT of roughly 15 per year; about 12 per 1000 corresponds to an NNT of roughly 84 per year. These are approximate values computed here from Cochrane's rounded numbers, not figures reported by the review, but the point does not depend on the decimals: the same intervention can look several times more or less efficient depending on who receives it. Cochrane notes that relative effects tend to be more stable across risk groups than absolute ones, which is why it recommends presenting absolute effects separately for each identifiable risk group.
Time frame. Cochrane's definition includes the words "in a given time frame", and its guidance on absolute effects asks authors to clarify the time period to which they apply. An NNT over one year and an NNT over five years are different numbers, because more events accumulate over longer follow-up. A 2026 methods paper in Nephrology Dialysis Transplantation by Chesnaye and colleagues, which explains how to calculate the NNT for both binary and time-to-event data, highlights "the importance of considering baseline risk, time, and confidence intervals when interpreting the NNT" (Chesnaye and colleagues, 2026).
Comparator. Cochrane stresses that the NNT is a comparative measure of effect, experimental versus a specific comparator, and "not a general property of a single intervention." An NNT against placebo and an NNT against another active treatment answer different questions, and neither can be carried over to the other.
What does an NNT not tell you?
It is an average, not a guarantee. In Cochrane's words, "NNT = 10 does not imply that one additional event will occur in each and every group of 10 people." The NNT describes an expected value across a population. It does not identify which individuals benefit.
It needs a confidence interval. Like any estimate from a sample, an NNT carries uncertainty. Douglas Altman set out how to attach confidence intervals to the NNT in a 1998 BMJ review (Altman, 1998). Cochrane explains one awkward feature: as a risk difference approaches zero, the NNT approaches infinity, so when the confidence interval for the risk difference includes zero, the interval for the NNT contains a discontinuity. Cochrane recommends labelling the two ends of such an interval separately, one as a number needed to treat for benefit and the other as one for harm. A single NNT reported without its interval hides how precise, or imprecise, the estimate is.
Harms use the same arithmetic. When a treatment increases the risk of an adverse outcome, the same calculation gives the number of people treated for one additional person to be harmed. Cochrane strongly recommends avoiding the term "number needed to harm" and its abbreviation, and instead using NNTB (number needed to treat for an additional beneficial outcome) and NNTH (number needed to treat for an additional harmful outcome). The paired terms make clear that both describe the same treated group. Chesnaye and colleagues also discuss the number needed to harm alongside a related measure, the likelihood to be helped or harmed, which sets benefit and harm against each other.
It should not be pooled naively. In systematic reviews, Cochrane warns that risk differences and NNTs should not be computed from the aggregated total numbers of participants and events across trials. Trials enrol people with different baseline risks, and simply adding them together can produce a misleading figure.
How should a reader use NNT when reading a study?
An NNT is most informative when it arrives with its context. A reader can ask:
- Compared with what? Placebo, usual care or another treatment.
- Over what period? The follow-up time the NNT refers to.
- In whom? The baseline risk of the group studied, and whether the reader's question concerns a similar group.
- How precise is it? Whether a confidence interval is reported, and whether that interval crosses from benefit into harm.
- What about harms? Whether an NNTH for important adverse outcomes is reported alongside the NNTB.
These questions are tools for reading evidence, not a basis for any individual decision.
Key takeaways
The number needed to treat is a useful translation of an absolute risk difference into people, and it keeps the size of an effect in view in a way a relative percentage cannot. It is also easy to misread. It changes with baseline risk and with follow-up time, it describes a comparison rather than a treatment, and it is an estimate that needs a confidence interval and a matching measure of harm. Read alongside related tools such as intention-to-treat analysis and an understanding of what a p-value actually tells you, it becomes one of the clearer ways to judge what a trial result means.
Readers interested in how study design and measurement shape the interpretation of evidence will find related work among Dr. Hessabi's peer-reviewed publications.