Number Needed to Treat: How Many People to Help One

Number needed to treat is the count of people who must take a treatment, for a defined period, for one of them to get the benefit being measured. A number needed to treat of 50 over five years means 50 people take it for five years and one avoids the event the study tracked, while the other 49 do not. The figure is the plain inverse of the absolute risk reduction, and that single move, turning a percentage into a head count, is what makes it one of the most honest ways to state what a treatment does. This is educational and not medical advice, so use it to read the evidence, then decide with your clinician.

I came to respect this number from both sides of it. My doctoral research at the Lund University Diabetes Centre is on the genetics of type 2 diabetes, where risk is the whole vocabulary, and in global drug development I watched one trial result get described in ways that sounded wildly different depending on which number led the sentence. Number needed to treat holds its shape under that pressure.

What is the number needed to treat?

Here is the quotable definition. The number needed to treat is how many patients you would treat, over a stated period, so that one additional patient avoids a defined bad outcome. It equals one divided by the absolute risk reduction, the plain difference in event rates between the treated and the comparison group.

Two words carry the weight. "Additional" means the number counts benefits that would not have happened anyway, not every good outcome in the treated group. "Defined" means the figure is meaningful only when it names the exact event and the time window. A 25 to prevent one heart attack over ten years is a different animal from a 25 to prevent one bruise over a month; a bare "25" with no event and no clock is not yet a fact.

Why it is one of the most honest numbers in medicine

Most ways of stating a benefit quietly flatter it. A relative reduction sounds the same whether the underlying risk was high or vanishingly small, because it hides how common the event was. Number needed to treat cannot be computed without the baseline, so it carries the baseline inside it.

It is also honest about the people who do not benefit. A number of 50 tells you in the same breath that 49 took on the cost, the routine, and the side effects without the measured benefit. Many excellent treatments have large numbers needed to treat, because we accept treating many to spare a few from something terrible. The figure simply puts that trade where you can see it.

A neutral worked example you can do in your head

Let me use invented round numbers so nothing rides on any real product. Picture 2,000 people in two equal groups of 1,000, followed for five years for one unwanted event. In the comparison group, 50 have the event, an absolute risk of 5 percent. In the treated group, 30 have it, an absolute risk of 3 percent. The absolute risk reduction is 2 percentage points, or 0.02 as a fraction.

Now invert it. One divided by 0.02 is 50, so the number needed to treat is 50: over five years, you treat 50 people like these for one to avoid the event. In relative terms, going from 50 events to 30 is a 40 percent reduction. Forty percent and a head count of fifty describe the identical result; one excites, the other tells you the scale of the work.

Change only the baseline and watch the number move. Suppose the event were ten times rarer, so the comparison group had 5 events and the treated group 3. The relative reduction is still 40 percent, but the absolute risk reduction is now 0.2 percentage points, and the number needed to treat balloons to 500. Same drug, same relative effect, ten times as many people treated for each one helped. That sensitivity is the figure telling you what the relative number conceals.

What about the number needed to harm?

Harm has its own version of the same idea. The number needed to harm is how many people you would treat, over a comparable window, for one additional person to suffer a specific side effect, again one over the absolute change in that harm. A treatment with a number needed to treat of 50 and a number needed to harm of 200 helps about four people for each one it harms that way. Setting the two counts side by side forces benefit and cost into the same unit, which reveals more than any list of percentages.

How to read a number needed to treat without being fooled

Two questions make the figure trustworthy. What exact outcome was prevented, and over how long? A small number for a surrogate marker, a lab value that moves, means much less than a slightly larger one for an event a patient would feel, like a stroke avoided. Pull the clock and the endpoint into view before you are impressed.

Then ask who was in the trial. The figure is anchored to the baseline risk of the people studied, so a number from a high-risk population overstates the benefit for a low-risk person, since the same relative effect over a smaller baseline yields a larger number needed to treat in real life. Treat the result as an estimate with a range around it, and give more weight to a benefit whose reported range stays comfortably finite than to one whose upper edge stretches toward thousands for a single gain.

Where this number earns its keep

The setting where I find it most clarifying is prevention in people who feel well, where baseline risk is often low and the number needed to treat can quietly climb into the hundreds. That does not make prevention pointless. It means the decision deserves the real count, weighed against the burden of taking something for years.

The number is not a verdict. A small one does not command you to treat, and a large one does not forbid it, because the weight you give to the outcome is yours and your clinician's to set together. What it does is keep the quiet majority who carry the effort without the reward in the same frame as the one who gains. Whenever a treatment is offered, ask for that count, the event, and the years it covers. A claim that can answer those is worth your attention. One that cannot is not finished.