Ask Mallard

Number needed to treat

The number needed to treat is one divided by the absolute risk reduction: how many patients must receive the treatment, for a given period, for one additional good outcome.

Why it is useful

It converts a probability into a count, which most people reason about more reliably. It also makes the comparison between treatments concrete: an NNT of 12 and an NNT of 300 are immediately different in a way that two risk ratios are not.

The same relative reduction at two baseline risks

Two arrays of one hundred people. With a higher baseline risk, twenty of a hundred have the outcome without treatment and ten with it: an absolute reduction of ten in a hundred, a relative reduction of fifty percent, and ten people treated to prevent one. With a lower baseline risk, two of a hundred without treatment and one with it: the same fifty percent relative reduction, but an absolute reduction of one in a hundred and a hundred people treated to prevent one.Higher baseline riskWithout treatment: 20 in 100With treatment: 10 in 100Absolute reduction: 10 in 100Relative reduction: 50%Number needed to treat: 10Lower baseline riskWithout treatment: 2 in 100With treatment: 1 in 100Absolute reduction: 1 in 100Relative reduction: 50%Number needed to treat: 100Dark: has the outcome even with treatment. Shaded: prevented by treatment. Empty: no outcome either way.The relative reduction is identical; what the treatment is worth to one person is not.
Worked arithmetic. The inputs are round illustrative numbers, not results from any study; the outputs are calculated from them.

What must accompany it

  • The TIME HORIZON. An NNT without a period is uninterpretable.
  • The BASELINE RISK it was calculated at, since it scales inversely with it.
  • The OUTCOME, precisely: the NNT for a composite is not the NNT for death.
  • The corresponding number needed to HARM for the important adverse effects, on the same horizon.
  • A confidence interval, with its awkwardness handled honestly rather than by omission.

The confidence interval problem

If the confidence interval for the risk difference crosses zero, the corresponding NNT interval is not a single range: it runs from a number needed to treat, out through infinity, and back from a number needed to harm. Reporting only the positive part hides that the treatment might do nothing or harm. The clean solution is to report the confidence interval on the RISK DIFFERENCE and give the NNT as a point estimate for communication.

It is not a threshold

There is no NNT above which a treatment is not worth giving. That depends on the severity of the outcome prevented, the burden and cost of the treatment, and the harms. An NNT of 500 is excellent for a cheap, safe intervention preventing death, and poor for an expensive one preventing mild symptoms.

What to watch out for

An NNT is meaningless without the time horizon and the baseline risk it was computed at, and both are routinely omitted: an NNT of 20 over five years is a different proposition from an NNT of 20 over six weeks. It also inherits every dependency of the absolute risk difference, so an NNT from a high-risk trial population does not apply to a low-risk patient. Confidence intervals for an NNT behave badly when the risk difference interval includes zero, producing a discontinuous range that cannot be printed as an ordinary interval.

Clinical example

A statin's NNT for preventing one cardiovascular event over five years is roughly 40 in a high-risk population and several hundred in a low-risk one, from the same relative effect. Quoting the first figure to the second patient is the error the baseline risk requirement exists to prevent.

Recent studies that used this
Searching PubMed…
Related Learn pages
Sources & further reading
  1. An assessment of clinically useful measures of the consequences of treatment (opens in a new tab) · N Engl J Med · doi 10.1056/NEJM198806303182605
  2. Using alternative statistical formats for presenting risks and risk reductions (opens in a new tab) · Cochrane Database Syst Rev · doi 10.1002/14651858.CD006776.pub2
Last source review: 2026-09-06
AI-generated educational content
Mallard uses reputable methodological and reporting sources, but this material may contain errors or omit important context. Verify important methodological decisions with the cited sources and, when appropriate, a qualified research methodologist or statistician.