By HTA Course · Published

Evidence synthesis

Number Needed to Treat (NNT): Formula, Calculation and Limits

The number needed to treat turns a trial result into a head count: how many patients must be treated for one of them to benefit. This guide explains the NNT formula, works through a calculation, covers the number needed to harm and confidence intervals, and shows why HTA bodies look beyond the NNT.

13-minute read Beginner level Biostatistics for HTA and HEOR

The number needed to treat (NNT) is the number of patients who need to receive a treatment, instead of the comparator, for one additional patient to benefit over a stated period. It is calculated as 1 divided by the absolute risk reduction, and a smaller NNT means a larger absolute effect.

The number needed to treat is the reciprocal of the absolute risk reduction (ARR): the risk of an unwanted event in the control group minus the risk in the treated group. An NNT of 20 means that, on average, 20 patients must be treated for the stated time to prevent one event that would have happened with the comparator.

The measure was introduced by Laupacis, Sackett and Roberts in the New England Journal of Medicine in 1988 and popularised by Cook and Sackett in the BMJ in 1995, who argued that it shows statistical and clinical significance in a single number. The short version is in our glossary entry on NNT.

Two points from the Cochrane Handbook are worth fixing in mind early. An NNT is a comparative measure: it describes one intervention against one specific comparator, as defined in the PICO question, and is not a property of the intervention alone. It is also an expected value: an NNT of 10 does not mean that exactly one extra event is prevented in every group of 10 patients.

NNT formula: how to calculate number needed to treat

The NNT formula is NNT = 1 ÷ ARR, where the ARR is the control event rate minus the experimental event rate, both written as proportions. If the risks are in percentages, use 100 ÷ ARR instead.

Patients treated per extra patient who benefits NNT
=
1 divided by the absolute risk reduction 1 ÷ (CER − EER)

CER is the control event rate and EER the experimental event rate, the labels used by the Centre for Evidence-Based Medicine (CEBM) in Oxford. The NNT calculation takes four steps:

  1. Extract the two risks. Events divided by patients in each arm, for the same outcome and the same follow-up.
  2. Calculate the ARR. CER − EER. The ARR is the same thing as the risk difference, with the sign ignored.
  3. Take the reciprocal. NNT = 1 ÷ ARR.
  4. Round up and label. Report a whole number together with the comparator, the outcome and the time frame.

Why the NNT is rounded up

By convention an NNT is always rounded up to the next whole number, never down: both the Cochrane Handbook and CEBM state the rule. An ARR of 0.06 gives 1 ÷ 0.06 = 16.7, reported as 17. The Handbook’s own examples follow the same rule: a risk difference of 0.12 gives an NNT of 9 (from 8.3) and one of 0.22 gives 5 (from 4.5). Rounding up avoids overstating the benefit.

Worked example: a large relative effect, a modest NNT

The same relative risk reduction can produce very different NNTs, because the NNT depends on how common the event is without treatment. The numbers below are hypothetical and rounded; they are not real trial data.

Suppose a new medicine is compared with standard care for one year in two populations. In high-risk patients, 20% have a heart attack on standard care and 15% on the new medicine. In low-risk patients the figures are 4% and 3%.

Measure High-risk patients Low-risk patients
Control event rate (CER) 20% (0.20) 4% (0.04)
Experimental event rate (EER) 15% (0.15) 3% (0.03)
Relative risk (EER ÷ CER) 0.75 0.75
Relative risk reduction (ARR ÷ CER) 25% 25%
Absolute risk reduction (CER − EER) 5 percentage points (0.05) 1 percentage point (0.01)
Events avoided per 1,000 treated 50 10
NNT for one year (1 ÷ ARR) 20 100

“Cuts heart attacks by a quarter” is true in both columns. Yet in the high-risk group 20 patients must be treated for a year to prevent one heart attack, and in the low-risk group 100. In each case the other 19 or 99 patients take the medicine without that benefit: some would never have had the event, and some have it anyway.

Cook and Sackett made the point with published data on blood pressure treatment. Over five years, the risk of stroke fell from 0.20 to 0.12 in moderate hypertension and from 0.015 to 0.009 in mild hypertension: a 40% relative risk reduction in both, but an NNT of 13 in one group and 167 in the other.

Relative versus absolute

A relative effect says how much the treatment changes the risk; the absolute effect says how many patients that change reaches. Relative measures such as the risk ratio, odds ratio and hazard ratio are compared in our guide to the hazard ratio. The NNT always comes from the absolute difference.

Number needed to harm (NNH)

The number needed to harm is the mirror image of the NNT: the number of patients who must be treated for one additional patient to experience a harmful outcome. It is 1 divided by the absolute risk increase, the risk of the adverse event with treatment minus the risk with the comparator.

Continuing the hypothetical example, suppose 6% of patients on the new medicine and 2% on standard care have a serious bleed within the year. The absolute risk increase is 0.04, so the NNH is 1 ÷ 0.04 = 25. Unlike the NNT, a larger NNH is better.

Put side by side, the two numbers give a rough benefit–harm picture. For every 100 high-risk patients treated for a year, about 5 heart attacks are prevented and about 4 extra bleeds are caused. An NNT of 20 against an NNH of 25 does not settle the question, because the two events are not equally serious; that weighting needs a common measure of health, discussed below.

The Cochrane Handbook strongly discourages the terms “number needed to harm” and “NNH”, because the phrase is easily misread as the number of people who will be harmed if given the intervention. It recommends NNTB (number needed to treat for an additional beneficial outcome) and NNTH (number needed to treat for an additional harmful outcome). NNH remains common in journals and search queries, so expect to meet both.

Confidence intervals for the number needed to treat

The confidence interval for an NNT is obtained by taking the reciprocals of the confidence limits of the ARR and swapping their order. Altman’s 1998 BMJ paper on the subject advises quoting a confidence interval whenever an NNT is reported.

In the hypothetical high-risk example, suppose the ARR is 0.05 with a 95% confidence interval of 0.01 to 0.09. The limits become 1 ÷ 0.09 = 11.1 and 1 ÷ 0.01 = 100, so the NNT is 20 (95% CI 11.1 to 100). A reciprocal stretches the uncertain end of the interval: the data are compatible with treating 12 patients or 100.

When the interval for the ARR crosses zero

If the result is not statistically significant, the interval for the ARR runs from a negative value (harm) to a positive one (benefit). An ARR of zero corresponds to an infinite NNT, so simply inverting the limits produces an interval that looks nonsensical.

Suppose the ARR is 0.05 with a 95% CI of −0.05 to 0.15. Inverting gives 20, with “limits” of −20 and 6.7, an interval that seems not to contain its own estimate. The explanation is that the interval passes through infinity, not through zero.

Altman’s NNTB and NNTH notation

Altman proposed labelling the two directions: NNTB when the treatment benefits and NNTH when it harms. The example becomes NNTB 20 (95% CI NNTH 20 to ∞ to NNTB 6.7). Read it as: the data are compatible with anything from one extra patient harmed per 20 treated, through no effect at all, to one extra patient helped per 6.7 treated.

The Cochrane Handbook uses the same labels and notes one further limitation: an interval calculated from the confidence limits of the summary effect estimate reflects uncertainty in the treatment effect but not in the baseline risk it is applied to.

Limits of the NNT and common mistakes

An NNT is only meaningful together with its context: the baseline risk of the patients, the length of follow-up, the comparator and the outcome. Most mistakes come from quoting or comparing NNTs with one of these stripped away.

1

Ignoring baseline risk

The worked example gave NNTs of 20 and 100 from the same relative effect. An NNT from a trial applies to patients at a similar risk to those in the trial.

2

Dropping the time frame

“NNT 20” is incomplete; “20 patients for one year” is a result. The Cochrane Handbook notes that studies with different follow-up will almost certainly have different comparator risks, and so different NNTs.

3

Comparing across trials

NNTs from trials with different comparators, outcomes, populations or durations are not like for like. A lower NNT against placebo does not show superiority over a medicine tested against active treatment.

4

Reading it as a guarantee

The NNT is an average across a group. It does not identify which patient benefits, and it is silent on how large the benefit is for that patient.

NNTs from a meta-analysis

A single NNT from pooled absolute risk differences can mislead, because the trials in a meta-analysis usually differ in baseline risk and follow-up. The Cochrane Handbook says risk differences are rarely appropriate for computing NNTs in systematic reviews, and Ebrahim reaches the same conclusion in a 2001 review of the pitfalls.

The recommended route is to pool a relative effect, then apply it to one or more assumed comparator risks (ACR): ARR = ACR × (1 − risk ratio). The Handbook’s example uses a risk ratio of 0.92 and an ACR of 0.3, giving 24 fewer events per 1,000 and an NNT of 42. With the hypothetical risk ratio of 0.75 used above, the NNT is 20 at a baseline risk of 20%, 40 at 10% and 100 at 4%: a range for different patient groups instead of one number. The same logic applies to results from a systematic literature review or a network meta-analysis, which is covered in Module 4: Evidence Synthesis.

NNTs from time-to-event data

For survival outcomes there is no single NNT: it can be calculated at any time point and will differ between them. Altman and Andersen’s 1999 BMJ paper gives two methods.

  1. From survival curves. Read the survival probabilities in both arms at the chosen time from the Kaplan–Meier curves; the NNT is 1 divided by their difference.
  2. From a hazard ratio. If only the hazard ratio (HR) and the control arm’s survival probability Sc at that time are known, survival with treatment is estimated as Sc raised to the power of the HR, and the NNT is 1 ÷ (ScHR − Sc).

A hypothetical illustration: with an HR of 0.70 and control survival of 85% at one year, survival with treatment is 0.850.70 = 0.892, the difference is 0.042 and the NNT is 24 (from 23.5). At a later time point where control survival is 50%, treated survival is 0.500.70 = 0.616 and the NNT is 9 (from 8.7). The hazard ratio is the same; the NNT is not. In one of Altman and Andersen’s own examples, a lung cancer trial, the NNT was 3.6 at six months and about 7 at three to four years.

How the number needed to treat is used in HTA and health economics

In health technology assessment the NNT is mainly a communication aid, not a decision metric. It makes an absolute effect easy to grasp, but it says nothing about how much health each patient gains, how long the gain lasts or what the treatment costs.

Three gaps explain why. First, an NNT treats every prevented event as equal: preventing one death and preventing one mild relapse can both be “NNT 20”. Second, it is tied to one outcome and one time point, while a decision has to weigh all benefits and harms over the relevant horizon. Third, it contains no costs.

Economic evaluation fills those gaps with different tools. NICE’s technology appraisal manual, for example, sets a reference case in which health effects are expressed in quality-adjusted life years (QALYs) and results are reported as incremental cost-effectiveness ratios. It also describes the net treatment effect as determined by the baseline risk and the relative effect of the technology, with relative effects from randomised trials applied to data on baseline risk. That is the same logic as calculating an NNT from a risk ratio, carried into a model.

The NNT still has a place. Clinical summaries often report absolute effects as events per 1,000 patients, the format Cochrane uses in its summary of findings tables, and an NNT is the same information turned upside down: 50 fewer events per 1,000 is an NNT of 20.

From NNT to cost per event avoided

A simple hypothetical shows both the link to economics and its limit. If the new medicine costs £1,000 more per patient per year than standard care and the NNT is 20, the extra drug cost is 20 × £1,000 = £20,000 per heart attack avoided. That is a cost-effectiveness ratio in natural units. It ignores the bleeds, the savings from fewer heart attacks and the value of avoiding one, so it cannot be set against a cost-per-QALY threshold.

Cost-effectiveness, thresholds and net benefit are taught in Module 7: Cost-Effectiveness and the ICER.

How to report and read an NNT: checklist

A usable NNT answers six questions. If any is missing from a paper, a slide or a value dossier, ask for it before using the number.

  1. Compared with what? Placebo, standard care or an active comparator.
  2. For which outcome? Including the definition or threshold if the outcome was dichotomised.
  3. Over what period? The follow-up or time point the risks refer to.
  4. At what baseline risk? The control event rate, and whether it matches the patients in question.
  5. How precise is it? A 95% confidence interval, in NNTB and NNTH notation if it includes no effect.
  6. What about harms? The matching NNH or NNTH for the main adverse events.

Key takeaways

  • NNT = 1 ÷ absolute risk reduction, rounded up to the next whole number.
  • The same relative risk reduction gives a small NNT at high baseline risk and a large one at low risk.
  • The number needed to harm is 1 ÷ absolute risk increase; Cochrane prefers NNTB and NNTH.
  • When the result is not significant, the confidence interval runs through infinity.
  • HTA decisions rest on absolute effects, QALYs and cost-effectiveness, not on the NNT alone.

Certificate Course in HTA

Learn to read clinical evidence like an HTA reviewer

Module 3 explains confidence intervals, risk ratios, hazard ratios, NNT and survival analysis; Module 4 shows how effects are pooled in meta-analysis and network meta-analysis; Module 7 turns them into cost-effectiveness results. The course presents both the industry and payer perspective without assuming prior HTA knowledge.

15 structured learning modules
79 interactive lessons
100% online and self-paced
Global HTA and market access perspective

Frequently asked questions

Number needed to treat basics

What is the number needed to treat?

The number needed to treat (NNT) is the number of patients who must receive a treatment, instead of the comparator, for one additional patient to benefit over a stated period. It is the reciprocal of the absolute risk reduction, so a smaller NNT means a larger absolute treatment effect.

How do you calculate number needed to treat?

Subtract the event rate in the treated group from the event rate in the control group to get the absolute risk reduction, then divide 1 by it and round up. If 20% of control patients and 15% of treated patients have the event, the reduction is 0.05 and the NNT is 20.

What is a good NNT?

There is no universal threshold. A lower NNT means more patients benefit, but whether a value is acceptable depends on the seriousness of the event prevented, the harms and cost of treatment, the duration of treatment and the baseline risk of the patients. An NNT should always be read with that context.

What is the difference between NNT and NNH?

The NNT counts patients treated for one extra patient to benefit; the number needed to harm (NNH) counts patients treated for one extra patient to have an adverse event. The NNH is 1 divided by the absolute risk increase. A low NNT and a high NNH are the favourable combination.

Can you compare NNTs from different trials?

Only with care. An NNT depends on the baseline risk of the patients, the length of follow-up, the comparator and the outcome definition. Trials that differ in any of these will give different NNTs even if the treatments work equally well, so a direct comparison can mislead.

Is the NNT used in health technology assessment?

Mainly as a way to communicate absolute effects. The NNT does not capture how much health each patient gains, how long the benefit lasts or what treatment costs, so HTA bodies that use economic evaluation rely on measures such as QALYs and incremental cost-effectiveness ratios for decisions.

Sources

  1. Laupacis A, Sackett DL, Roberts RS. An assessment of clinically useful measures of the consequences of treatment. New England Journal of Medicine 318(26):1728–1733, 1988 (bibliographic record).
  2. Cook RJ, Sackett DL. The number needed to treat: a clinically useful measure of treatment effect. BMJ 310(6977):452–454, 1995.
  3. Altman DG. Confidence intervals for the number needed to treat. BMJ 317(7168):1309–1312, 1998.
  4. Altman DG, Andersen PK. Calculating the number needed to treat for trials where the outcome is time to an event. BMJ 319(7223):1492–1495, 1999.
  5. Barrowman NJ. Missing the point (estimate)? Confidence intervals for the number needed to treat. CMAJ 166(13):1676–1677, 2002.
  6. Schünemann HJ, Vist GE, Higgins JPT, et al. Chapter 15: Interpreting results and drawing conclusions. Cochrane Handbook for Systematic Reviews of Interventions, version 6.5, 2024.
  7. Centre for Evidence-Based Medicine, University of Oxford: Number needed to treat (NNT).
  8. Ebrahim S. The use of numbers needed to treat derived from systematic reviews and meta-analysis: caveats and pitfalls. Evaluation & the Health Professions 24(2):152–164, 2001.
  9. NICE. Technology appraisal and highly specialised technologies guidance: the manual (PMG36), published 31 January 2022, last updated 31 March 2026, sections 4.2, 4.6 and 4.9.

Not ready to enrol yet?

Get the full course syllabus by email

  • All 15 modules, from evidence and statistics to modelling and reimbursement
  • Prices: 3 months €169, 6 months €289, 12 months €490
  • A short note to forward if your employer pays for training (VAT invoice available)

One email with the syllabus. See our privacy policy.

Article rating

Rate this article

Be the first to rate this article

Click a star to rate

Last updated: 18/11/2026