NNT, NNH, and NNS: benefit and harm in whole people

Worth the Test · Written by an emergency physician · Last reviewed September 2026

"This drug cuts heart-attack risk by a third" and "104 people must take this drug for five years for one of them to avoid a heart attack" can describe the same trial. The first phrasing sells; the second informs. The number needed to treat — NNT — is medicine's most effective honesty device: it converts a treatment effect from a percentage change, which minds inflate, into a count of whole people, which minds can actually picture. This guide covers the NNT, its shadow the NNH, and the screening cousin NNS that gets confused with both.

From risk difference to a count of people

The machinery is one division. If untreated people suffer some outcome at rate x and treated people at a lower rate y, the absolute risk reduction is ARR = x − y, and

NNT = 1 ÷ ARR.

Treat a group whose five-year heart-attack risk is 2% with something that lowers it to 1%: the ARR is one percentage point, so NNT = 1 ÷ 0.01 = 100 people treated for five years per heart attack prevented. The measure was introduced by Laupacis, Sackett, and Roberts in 1988 precisely because the alternative framing misleads: that same treatment boasts a "50% relative risk reduction," and a 50% reduction sounds like it helps every other person who takes it. It helps one in a hundred. Relative risk reduction describes what happens to the risk; the NNT describes what happens to people.

The same division explains why one drug can carry wildly different NNTs. A treatment's relative effect is often roughly stable across populations, but the baseline risk it multiplies is not — halve a 20% risk and NNT is 10; halve a 0.2% risk and NNT is 1,000. Same pill, hundredfold difference, and neither number is "the" NNT of the drug. An NNT is always a statement about a treatment in a population with a particular baseline risk, over a particular time. Quoted without the population and the horizon, it is decoration.

Real NNTs, from the trials

Worked from published meta-analyses (each figure's source is in the references):

• Statins after a heart attack or stroke (secondary prevention): about 1 death prevented per 83 people treated for five years — from the Cholesterol Treatment Trialists' pooled analysis of fourteen randomized trials.
• Statins in people without heart disease (primary prevention): about 1 heart attack prevented per 104 treated for five years — from the Cochrane review. Same drug class as above; the lower-risk population stretches the NNT.
• Antibiotics for a child's ear infection: about 1 child in 7 to 20 with less pain at days two to twelve — a reminder that NNTs apply to symptom relief, not just death.

Notice what these numbers do to intuition. Statins are among the most prescribed drugs on earth, backed by unusually strong evidence — and even in the highest-risk group, 82 of 83 people take them for five years without their own death being the one prevented. That is not an argument against statins; it is what effective preventive medicine actually looks like when counted in whole people. If that surprises you, the number is doing its job.

NNH: the same division, run on harms

Treatments also cause things, and the identical arithmetic on an absolute risk increase yields the number needed to harm. For people who have already had a heart attack or stroke, about two years of daily aspirin prevents roughly one serious vascular event (a heart attack, stroke, or vascular death) per 50 people treated, and causes roughly one major bleed per 400 — figures computed by theNNT.com from meta-analyses by the Antithrombotic Trialists' Collaboration. Antibiotics for ear infections: benefit in 1 per 7–20, vomiting, diarrhea, or rash in about 1 per 14.

The harm half of a pair is often the less certain one. A Cochrane review of warfarin for people with atrial fibrillation and no history of stroke or transient ischemic attack estimated about 25 strokes prevented for every 1,000 patients treated for a year — roughly one per 40 — but called its bleeding estimates imprecise: hemorrhage was not significantly increased, and the confidence intervals were wide. A harm that has not been measured precisely has not been shown to be absent.

An NNT next to an NNH looks like a ready-made verdict: helped-per-harmed. Resist the reflex, for two reasons. First, the events differ in weight — a death prevented and a transient rash are not exchangeable units, and for statins the commonly quoted harm (muscle symptoms in roughly 1 in 10, an estimate from theNNT.com that runs well above the myopathy rates seen in blinded trials — the figure itself is contested) is mostly mild and reversible, while the benefit counted is death. Second, both numbers inherit all the population- and horizon-dependence above. The honest use of an NNT/NNH pair is not a ratio but a sentence: out of every hundred people like this who take this for this long, about one avoids X and about ten experience Y — and then you weigh X against Y with your own values. The calculator's outcome view turns that sentence into a picture. It shows the expected number helped and the expected number harmed as separate effects that can overlap — one person can be both — plus the treated people with neither modeled effect. "Neither" is usually the largest group, but not always: with an NNH of 1, everyone treated is harmed and nobody is left in it.

NNS: the screening cousin, and why it dwarfs both

Screening programs quote a third number: the number needed to screen — how many people must be screened (not treated) for one to avoid the outcome. For low-dose CT lung screening, about 320 screened per lung-cancer death prevented; for mammography, the Cochrane review's contested estimate runs around 2,000 invited per breast-cancer death avoided. Other bodies read the mammography evidence more favorably; the screening-harms page sets the Cochrane reading beside the U.S. Preventive Services Task Force's. An NNS is usually far larger than a treatment NNT, because most people screened do not have the disease the program is looking for.

A trial NNS summarizes a screening program’s population, invitation or attendance, repeated testing, follow-up care and outcome horizon. The calculator instead reports screens per modeled benefit: everyone screened divided by the expected benefit among treated true positives, for one modeled round. Both are useful when properly labeled, but they are not interchangeable. The lung-CT example shows why entering an NNS of 320 into the treatment box creates a large denominator error. One cannot infer a trial’s treatment NNT from screening NNS alone.

Using the numbers without being used by them

Three habits keep these numbers honest. Ask per how long — an NNT without a time horizon is meaningless, and the NNT cannot generally be extrapolated by doubling the follow-up. Ask in whom — find the baseline risk the figure assumes and check it resembles the person in front of you. And ask NNT of what, against NNH of what — insist on knowing both events before weighing them. None of this requires more math than the one division this page opened with; it requires only refusing to let a single glossy number stand in for the four or five unglamorous ones underneath it.

Correction, September 2026. An unsupported figure of one major bleed per 25 people taking anticoagulants for atrial fibrillation was removed, and the stroke benefit now matches the Cochrane review: about 25 strokes prevented per 1,000 patients per year, not one per 25. The calculator's helped and harmed counts are now described as effects that can overlap. The corrections log lists every change.