# How to Calculate Number Needed to Treat (NNT) With a Confidence Interval

The number needed to treat (NNT) is the average number of patients who must receive a treatment for one additional patient to avoid an unfavorable outcome. It comes from the same 2x2 table you already use for risk ratios, but it expresses the result on a scale clinicians and reviewers can act on: patients per event prevented. The measure was introduced by Laupacis, Sackett and Roberts as a way to translate trial results into a clinically interpretable quantity [1].

A point estimate alone is not enough for a lab report or a manuscript. If the confidence interval for the risk difference crosses zero, the NNT confidence interval is not a tidy range: it splits into a number needed to treat for benefit (NNTB) and a number needed to treat for harm (NNTH), with an infinitely large value in between [3][4]. This article covers the NNT formula, the interval calculation, and the reporting conventions you should follow.

## Quick Answer

- NNT is the reciprocal of the absolute risk reduction (ARR): $NNT = \frac{1}{ARR}$ [2].
- ARR is the control event risk minus the treated event risk: $ARR = p_{control} - p_{treated}$.
- The confidence interval for NNT is obtained by inverting the confidence interval for the risk difference, not by inverting the endpoints of a relative effect [3].
- If the risk difference interval includes zero, report NNTB and NNTH separately, with infinity at the boundary [3][4].
- Round NNT values up to the next whole number by convention [4].

## Step 1: Build the 2x2 Table and Get the Risks

Everything starts with counts, not percentages. For a two-arm study with a binary outcome, tabulate events and non-events in each arm. Let $a$ be events in the treated group, $n_t$ the treated group size, $c$ events in the control group, and $n_c$ the control group size.

The event risk in each arm is a simple proportion:

$$p_{treated} = \frac{a}{n_t}, \qquad p_{control} = \frac{c}{n_c}$$

Keep the risks as proportions between 0 and 1 for the arithmetic, then convert to percentages when you write them up. Rounding early is a common source of drift, so carry at least four decimal places through the intermediate steps.

## Step 2: Compute the Absolute Risk Reduction

The ARR is the difference between the two risks:

$$ARR = p_{control} - p_{treated}$$

A positive ARR means the treatment lowered the event risk. A negative ARR means the treatment group had more events, which is a different clinical message even if the magnitude is identical.

This is where NNT separates from relative measures. The relative risk reduction (RRR) is the proportional drop, $RRR = (p_{control} - p_{treated}) / p_{control}$, and it is often assumed to stay roughly constant across baseline risks [2]. The ARR does not. The same RRR produces very different ARRs, and therefore very different NNTs, depending on how sick the population is. The difference between relative and absolute risk reduction is the most important thing to understand before you report an NNT.

## Step 3: Invert the ARR to Get NNT

The NNT formula is a reciprocal:

$$NNT = \frac{1}{ARR} = \frac{1}{p_{control} - p_{treated}}$$

If the ARR is 0.10, the NNT is 10. Read that as: for every 10 patients treated, one additional event is avoided compared with control, on average, over the same follow-up period as the study. The NNT is a property of the comparison, the outcome, and the time horizon. Change any of those and the number changes.

Because NNT is a count of patients, it is reported as a whole number. The Cochrane Handbook notes that NNTs are rounded up to the next whole number by convention [4]. An NNT of 9.2 becomes 10. Rounding up is conservative: it does not claim more efficiency than the data support.

## Step 4: Get the Confidence Interval for the Risk Difference

The interval for NNT comes from the interval for the risk difference (RD), which is the same quantity as the ARR when it is computed as control risk minus treated risk. For a large-sample Wald interval:

$$RD \pm z_{1-\alpha/2} \sqrt{\frac{p_{treated}(1-p_{treated})}{n_t} + \frac{p_{control}(1-p_{control})}{n_c}}$$

The term under the square root is the squared standard error of the difference of two independent proportions. For a 95% interval, $z_{1-\alpha/2}$ is 1.96.

The Wald interval is the one most people compute by hand, but it behaves poorly when events are rare or groups are small. Newcombe compared eleven interval methods for the difference between independent proportions and showed that the choice of method changes coverage [5]. For small samples or extreme proportions, use a score-based or exact method instead of Wald, and say which method you used. The Cochrane Handbook covers effect measure computation and the assumptions behind each approach [6].

## Step 5: Invert the Interval Endpoints

Once you have the RD interval, invert each endpoint to get the NNT interval [3]. If both endpoints are on the same side of zero, the inversion is straightforward and the interval is finite.

The direction matters. If the RD interval runs from 0.0222 to 0.1778, the NNT interval runs from $1/0.1778$ to $1/0.0222$, that is, from about 5.62 to about 45.01. The larger risk difference gives the smaller NNT, so the endpoints swap order when you invert. This is the step people most often get backwards.

If the RD interval includes zero, the inversion breaks. As the risk difference approaches zero from the positive side, NNT approaches positive infinity. As it approaches zero from the negative side, NNT approaches negative infinity. The interval is therefore discontinuous, and a single finite range cannot describe it [4].

## Worked Example

The following data are hypothetical and used only to illustrate the arithmetic. Suppose a trial randomizes 200 patients to treatment and 200 to control. In the treatment arm, 30 patients have the event. In the control arm, 50 do.

| Arm | Events | Total | Risk |
|---|---|---|---|
| Treatment | 30 | 200 | 0.15 |
| Control | 50 | 200 | 0.25 |

The risks are $p_{treated} = 0.15$ and $p_{control} = 0.25$. The absolute risk reduction is:

$$ARR = 0.25 - 0.15 = 0.10$$

The point estimate of NNT is:

$$NNT = \frac{1}{0.10} = 10$$

For the interval, the Wald standard error of the risk difference is 0.03969. The 95% confidence interval for the risk difference runs from 0.0222 to 0.1778. Inverting those endpoints gives an NNT confidence interval of 5.62 to 45.01. Because both endpoints of the risk difference interval are positive, the NNT interval is finite and describes benefit throughout.

For context, the same table gives a risk ratio of 0.60 (95% CI 0.399 to 0.902), an odds ratio of 0.529 (95% CI 0.320 to 0.876), and a relative risk reduction of 40%. A chi-square test without continuity correction gives 6.25 with p = 0.0124. Notice how much wider the NNT interval looks than the risk ratio interval. That width is real, not an artifact: NNT is a reciprocal, so it amplifies uncertainty near small risk differences.

You can reproduce these numbers with the [Odds Ratio, Relative Risk and NNT Calculator](/tools/odds-ratio-calculator) if you want to check your hand calculation.

### A Non-Significant Variant

Change the treatment events from 30 to 40, keeping everything else the same. The risks are now 0.20 and 0.25, so the risk difference is 0.05 with a 95% confidence interval from -0.0317 to 0.1317. The interval crosses zero.

The point estimate is NNTB = 20. The interval cannot be written as a single range. Following the NNTB/NNTH notation [3][4], you would report: NNTB 20, 95% CI NNTB 7.59 to infinity to NNTH 31.55. The "to infinity to" phrasing is awkward but correct. It tells the reader that the data are compatible with a benefit as large as one event prevented per 7.59 patients, with no effect at all, and with harm as large as one extra event per 31.55 patients.

## Step 6: Adjust NNT for a Different Baseline Risk

An NNT of 10 applies to a population with a 25% control event risk. Your patient or your study population may have a different baseline risk. If the relative risk reduction is assumed constant, you can extrapolate [2].

Take the RRR of 40% from the worked example and apply it to other baseline risks:

| Control risk | ARR at RRR 0.40 | NNT |
|---|---|---|
| 25% | 0.10 | 10 |
| 5% | 0.02 | 50 |
| 1% | 0.004 | 250 |

The treatment effect is identical in relative terms across all three rows. The NNT is not. This table is the clearest argument for reporting the baseline risk alongside any NNT, and for being cautious when applying a trial NNT to a population with a very different event rate.

## Common Mistakes

- **Inverting the relative risk or odds ratio instead of the ARR.** NNT is the reciprocal of the absolute risk reduction, not of RR or OR [2]. Inverting an OR gives a number with no clinical meaning.
- **Inverting the NNT confidence interval endpoints without swapping them.** The larger risk difference maps to the smaller NNT. Reverse the order when you invert.
- **Reporting a single finite NNT interval when the risk difference interval crosses zero.** That interval is discontinuous and must be reported as NNTB and NNTH with infinity at the boundary [3][4].
- **Rounding NNT down.** Round up to the next whole number [4]. Rounding down overstates the efficiency of treatment.
- **Dropping the time horizon.** An NNT without a follow-up period is incomplete. State the outcome and the duration.
- **Using the Wald interval for rare events or small samples.** Coverage degrades. Use a score or exact method and name it [5].
- **Treating NNT as a fixed property of a drug.** It depends on baseline risk, outcome definition, and follow-up. It is an estimate from a specific comparison.

## Limitations

NNT is undefined when the ARR is exactly zero, and unstable when the ARR is close to zero. Small changes in event counts can move the estimate by an order of magnitude, which is why the confidence interval is often very wide even when the relative effect is precisely estimated.

The constant-RRR extrapolation in Step 6 is an assumption, not a fact. It fails when the treatment has a threshold effect, when competing risks differ across populations, or when baseline risk is so low that the absolute benefit becomes negligible.

NNT also compresses information about harm. A single number cannot represent both benefit and harm for the same patient. When a treatment has important adverse effects, report the number needed to harm alongside the NNT, computed the same way from the adverse event risks, and be explicit about which outcome each number refers to. The two numbers are not directly comparable unless the outcomes are of similar severity and occur over the same time frame.

Finally, NNT is a group average. It does not identify which individual patient will benefit. A patient whose baseline risk is far from the trial average may have a very different absolute benefit, and the NNT from the trial may not transfer.

## Frequently Asked Questions

### What is the difference between NNT and absolute risk reduction?

ARR is the arithmetic difference in event risks between two groups, expressed as a proportion. NNT is the reciprocal of that difference, expressed as a count of patients. They carry the same information, but NNT is easier to communicate to a clinical audience because it maps onto a decision about how many patients to treat.

### Can NNT be negative?

Yes. A negative NNT means the treatment group had more events than the control group, so the "treatment" is associated with harm. Many authors avoid negative values by reporting the number needed to harm instead, which is the reciprocal of the absolute risk increase. State clearly which direction you are reporting.

### Why does the NNT confidence interval include infinity?

When the confidence interval for the risk difference includes zero, the data are compatible with no effect. Since NNT is the reciprocal of the risk difference, a risk difference of zero maps to an infinite NNT. The interval therefore has a gap, with NNTB on one side and NNTH on the other [3][4].

### Should I use the Wald interval for NNT?

The Wald interval is fine for moderate to large samples with event risks away from 0 and 1. For small trials, rare events, or proportions near the boundaries, use a score-based or exact method for the risk difference and invert that interval instead [5]. Report which method you used so readers can assess it.

### How do I report NNT in a manuscript?

Give the point estimate, the confidence interval, the outcome, and the follow-up period. If the risk difference interval crosses zero, use the NNTB/NNTH notation with infinity at the boundary [3][4]. Round to whole numbers, rounding up [4]. Include the baseline risk in each group so readers can judge whether the NNT applies to their population.

## References

1. [Laupacis A, Sackett DL, Roberts RS. An assessment of clinically useful measures of the consequences of treatment. New England Journal of Medicine, 1988](https://doi.org/10.1056/NEJM198806303182605)
2. [Cook RJ, Sackett DL. The number needed to treat: a clinically useful measure of treatment effect. BMJ, 1995 (PubMed)](https://pubmed.ncbi.nlm.nih.gov/7873954/)
3. [Altman DG. Confidence intervals for the number needed to treat. BMJ, 1998 (PubMed)](https://pubmed.ncbi.nlm.nih.gov/9804726/)
4. [Cochrane Handbook for Systematic Reviews of Interventions, Chapter 15: Interpreting results and drawing conclusions](https://training.cochrane.org/handbook/current/chapter-15)
5. [Newcombe RG. Interval estimation for the difference between independent proportions: comparison of eleven methods. Statistics in Medicine, 1998 (PubMed)](https://pubmed.ncbi.nlm.nih.gov/9595617/)
6. [Cochrane Handbook for Systematic Reviews of Interventions, Chapter 6: Choosing effect measures and computing estimates of effect](https://training.cochrane.org/handbook/current/chapter-06)

## Related Articles

- [Relative Risk vs Odds Ratio Interpretation](/knowledge/bioinformatics/how-to-calculate-and-interpret-relative-risk-and-odds-ratios-in-cohort-and-case-control-studies-a-bi)
- [Measures of Association in Veterinary Epidemiology: Risk and Odds Ratios](/knowledge/veterinary-medicine/veterinary-epidemiology/measures-association-veterinary-epidemiology-risk-odds-ratios)
- [Effect Sizes for Categorical Data](/knowledge/bioinformatics/effect-sizes-for-categorical-data-cramer-s-v-phi-and-odds-ratio-how-to-choose-and-report-them)
- [Effect Size Explained](/blog/research-skills/effect-size-explained-how-to-measure-the-magnitude-of-a-finding-without-getting-lost-in-p-values)
- [Binomial Effect Size Display: A Tool for Interpreting Effect Sizes](/blog/guides/binomial-effect-size-display-a-tool-for-interpreting-effect-sizes)
- [Statistical Significance vs. Practical Significance](/blog/research-skills/statistical-significance-vs-practical-significance-what-really-matters-in-research-results)