The number needed to treat is the average number of patients who must receive a treatment, rather than the comparison or control condition, for one additional person to benefit over a defined follow-up period. It translates a study's statistical effect into a plain count that clinicians and patients can picture: if the number needed to treat is 10, then treating 10 people produces one extra good outcome that would not have happened otherwise. It is derived directly from the absolute risk reduction, the simple difference in event rates between the two groups, which makes it one of the most clinically tangible ways to express how much a treatment actually does.
How it is calculated from the absolute risk reduction
The number needed to treat is the reciprocal of the absolute risk reduction. You first compute the control event rate, the proportion of people in the comparison group who had the outcome, and the treatment event rate, the proportion in the treated group who had it. The absolute risk reduction is the control event rate minus the treatment event rate. Dividing 1 by that difference gives the number needed to treat. Because it is a reciprocal, a tiny absolute difference produces a large number, and a big absolute difference produces a small number; the arithmetic is the entire reason a treatment with a modest absolute effect can still carry an impressive-sounding relative claim.
- Control event rate is the events in the comparison group divided by the total in that group.
- Treatment event rate is the events in the treated group divided by the total in that group.
- Absolute risk reduction is the control event rate minus the treatment event rate.
- Number needed to treat is 1 divided by the absolute risk reduction, rounded up to the next whole person because you cannot treat a fraction of a patient.
Keep the units consistent. If the absolute risk reduction is expressed as a proportion such as 0.05, the number needed to treat is 1 divided by 0.05, or 20. If you prefer percentages, divide 100 by the percentage-point difference instead. Both routes give the same answer as long as you do not mix a proportion with a percentage in the same division.
A worked numeric example
Suppose a trial of a preventive medication enrols 1000 people on treatment and 1000 on a placebo. Over two years, 80 people in the placebo group have a stroke and 50 people in the treatment group have one. The control event rate is 80 divided by 1000, or 0.08; the treatment event rate is 50 divided by 1000, or 0.05. The absolute risk reduction is 0.08 minus 0.05, which equals 0.03. The number needed to treat is 1 divided by 0.03, which is 33.3, so you round up and report that you must treat 34 people for two years to prevent one additional stroke.
Notice how differently the same data can be framed. The same trial supports a relative risk reduction of roughly 38 percent, because 0.03 is about 38 percent of the 0.08 baseline. The relative figure sounds dramatic, yet the absolute picture, one stroke prevented for every 34 people treated, is far more sober. This gap between the relative and the absolute story is exactly why thoughtful reporting pairs them. If you want to see the relative side worked through on its own, the companion explainer on how relative risk reduction is computed walks through the same kind of arithmetic, and you can check any of these figures instantly with the number needed to treat calculator.
Interpreting low versus high values
The number needed to treat is read in the direction of smaller-is-better. A low number needed to treat means few patients have to be treated for one to benefit, which signals a powerful and efficient intervention. A short course of antibiotics that cures almost everyone might have a number needed to treat near 2 or 3. A high number needed to treat means many patients must be treated before a single one gains, which is common for preventive therapies given to large, mostly healthy populations where the baseline event rate is already low.
Crucially, there is no universal threshold for a "good" number needed to treat. The judgement depends on the baseline risk, the seriousness of the outcome being prevented, the cost and burden of the treatment, and the harms it might cause. A number needed to treat of 100 can be entirely worthwhile when the prevented outcome is death and the treatment is a cheap daily tablet with few side effects, while a number needed to treat of 5 can be unattractive if the outcome is mild and the treatment is expensive or toxic. Always interpret it against the time horizon too, because a number needed to treat over five years is not comparable to one measured over six months.
The number needed to harm
The mirror image of benefit is the number needed to harm, the average number of patients who must receive a treatment for one additional person to experience a specified adverse outcome. It is calculated the same way, as the reciprocal of the absolute risk increase in harm between the treated and the comparison groups. If a drug causes a serious bleed in 1 extra person for every 200 treated, the absolute risk increase is 0.005 and the number needed to harm is 200.
Reading the number needed to treat and the number needed to harm side by side is the heart of a balanced benefit-risk assessment. If a treatment has a number needed to treat of 25 to prevent one heart attack but a number needed to harm of 400 for a serious bleed, the benefit clearly outweighs the harm at the population level. When the two numbers are close, the decision becomes far more finely balanced and depends heavily on how patients value the specific benefit against the specific harm. Understanding the underlying ratio measures helps here; the guide on the difference between the odds ratio and the risk ratio explains the relative metrics that sit behind these absolute counts.
Confidence intervals on the number needed to treat
A point estimate is never the whole story, because every trial result carries sampling uncertainty. The honest way to express that uncertainty is a confidence interval, and the standard approach is to build the interval on the absolute risk reduction first, then take reciprocals of its two ends. If the absolute risk reduction is 0.03 with a confidence interval from 0.01 to 0.05, the number needed to treat is 33 with an interval running from 20, the reciprocal of 0.05, to 100, the reciprocal of 0.01. The wider the interval, the less certain you are about how many patients truly need treating.
A well-known wrinkle appears when the result is not statistically significant. If the confidence interval for the absolute risk reduction crosses zero, the treatment might raise or lower risk, and the reciprocal interval splits into a confusing region that spans a number needed to treat and a number needed to harm with an apparent gap through infinity. The accepted convention, often credited to the statistician Douglas Altman, is to report the interval as running from a number needed to treat at one end, through infinity, to a number needed to harm at the other, rather than pretending it is a tidy single range. Always report the interval, never the point estimate alone, because an unqualified number needed to treat hides exactly how fragile the estimate may be.
How it connects to meta-analysis pooling
In a systematic review, the number needed to treat is rarely pooled directly. The preferred practice is to combine studies on a relative measure, usually the risk ratio or the odds ratio, on the log scale, because relative effects tend to be more consistent across populations with different baseline risks. The pooled relative effect is then applied to a chosen, representative baseline risk to back-calculate a single, clinically meaningful number needed to treat for the review. This two-step approach keeps the statistical pooling stable while still delivering the absolute number that decision-makers want.
- Pool the relative effect across studies first, because absolute differences travel poorly when baseline risks vary widely.
- Choose a representative baseline risk from the control arms or from a relevant external population.
- Apply the pooled relative effect to that baseline to derive an absolute risk reduction, then take its reciprocal for the number needed to treat.
- Present a range of baseline risks when populations differ, since the same relative effect yields a smaller number needed to treat in high-risk groups and a larger one in low-risk groups.
This is also why two reviews of the same drug can report different numbers needed to treat without contradicting each other: they may have assumed different baseline risks. When you set up the underlying synthesis, the overview of how a meta-analysis combines studies explains the pooling logic, you can compute the relative effects with the risk ratio calculator, run the synthesis in the pooled effect calculator, and turn the combined counts into a published figure with the forest plot builder. Reported with its confidence interval and an explicit baseline risk, the number needed to treat becomes one of the clearest summaries a review can offer.