Number Needed to Treat

Number Needed to Treat: Concept Architecture

Number needed to treat (NNT) expresses an absolute difference between an intervention and a comparator as the expected number of people who would need the intervention for one additional person to experience or avoid a defined event. It is tied to a specific outcome, population, and follow-up period. This page derives it from risks, works through benefit and uncertainty, and shows why time-to-event and competing-risk data require careful estimation.

Start with two comparable absolute risks

For a harmful event such as stroke, define $p_C(t)$ as its probability by a specified time $t$ under a comparator and $p_T(t)$ as its probability under treatment in the same target population. A positive absolute risk reduction is $\mathrm{ARR}(t)=p_C(t)-p_T(t)>0$. The unrounded number needed to treat for an additional beneficial outcome is its reciprocal:

$$\mathrm{NNTB}(t)=\frac{1}{p_C(t)-p_T(t)},\qquad p_C(t)>p_T(t).$$

Risks enter the denominator as proportions, not percentages: an eight-percentage-point reduction is $0.08$, and its reciprocal is 12.5, not $1/8$. By reporting convention the final NNT is rounded up to a whole person, giving 13 in this example. State the unrounded calculation when it helps audit the estimate. When the treatment instead increases the probability of a harmful event, label the reciprocal of $p_T-p_C$ as the number needed to treat for an additional harmful outcome (NNTH), rather than describing a negative NNTB. For a desirable event, such as recovery, the benefit difference is $p_T-p_C$; always define the event and direction before taking a reciprocal.

An NNT is a comparative expectation, not a literal claim that exactly one of every group of 13 people will benefit. It does not identify which person benefits, nor is it an intrinsic property of a drug without the comparator and context. Treatment allocation, adherence, crossover, competing mortality, and the analysis population affect the risks that underlie the reported comparison.

Worked example with a fixed horizon

Suppose an illustrative randomized comparison finds that 20% of comparator participants and 12% of treated participants experience the specified adverse event over two years. The ARR is $0.20-0.12=0.08$, or eight fewer events per 100 treated people on average under the comparison. The unrounded NNTB is $1/0.08=12.5$, reported as NNTB 13 over two years for that event, population, and comparator.

The risk ratio here is $0.12/0.20=0.60$, corresponding to a 40% relative reduction. That relative figure does not determine NNT without a baseline risk. If the same risk ratio of 0.60 applied to a lower-risk population with comparator risk 0.05, the modeled treatment risk would be 0.03, ARR would be 0.02, and NNTB would be 50 over the same horizon. Transporting the relative effect across populations is an assumption to assess, not a fact established by the first trial.

Interpret uncertainty on the risk-difference scale

A confidence interval for the absolute risk difference should accompany the NNT when available, because reciprocals become unstable near zero. If an illustrative 95% confidence interval for ARR were 0.02 to 0.14, its reciprocal limits would be approximately 7.14 and 50: under whole-person rounding, the corresponding beneficial NNT limits are about 8 to 50. These values are for the stated hypothetical interval; they do not follow from the four risks in the worked example without sample sizes and an inferential analysis.

If an ARR interval crosses zero, for example from $-0.02$ to $0.14$, the corresponding NNT scale passes through infinity. The interval permits both benefit (from about NNTB 8 toward infinity) and harm (from infinity toward NNTH 50). It must not be printed as an ordinary finite interval “8 to 50” because that incorrectly suggests one direction of effect throughout. Report the underlying risk difference and its interval prominently, label NNTB and NNTH limits by direction, and explain the discontinuity. At an exactly zero risk difference, no finite NNT is defined.

The rounded NNT is an aid to communication, not a precise biological constant. A difference between NNT 12 and 13 may reflect rounding and uncertainty rather than a meaningful clinical contrast. Clinical importance depends on event severity, baseline risk, treatment burden, alternative outcomes, preferences, and the whole evidence base—not on one reciprocal alone.

Use time-specific risks for survival outcomes

When follow-up differs across participants or an event is right-censored, raw observed event counts divided by initial group sizes generally do not estimate risk by a common time. Choose a time $t$ and estimate each group's cumulative event probability at that time with methods appropriate to censoring. Without competing events and under suitable assumptions, if $S_C(t)$ and $S_T(t)$ are event-free survival probabilities, the adverse-event risk reduction is $[1-S_C(t)]-[1-S_T(t)]=S_T(t)-S_C(t)$.

$$\mathrm{NNTB}(t)=\frac{1}{S_T(t)-S_C(t)},\qquad S_T(t)>S_C(t).$$

For example, an NNT at one year may differ from the NNT at five years because cumulative risks can change with time. A hazard ratio alone is not an NNT; conversion requires a baseline survival function, an explicit time, and a justified treatment-effect model. Do not use the difference in incidence rates per person-year as though it were a difference in cumulative event probabilities. If death or another event prevents the outcome of interest, use the appropriate cumulative incidence for the competing-risk question; treating a competing event as ordinary censoring can distort the absolute risk used for NNT.

Keep benefit, harm, and economic value distinct

The same intervention can prevent one event and cause another. Report separate absolute risks, time horizons, and NNTB or NNTH values for clearly defined outcomes. Comparing their numerical sizes alone cannot resolve the decision: outcomes differ in severity, timing, duration, uncertainty, and patient values. An outcome with different baseline risks in clinically identifiable subgroups can have different NNTs even under an assumed common relative treatment effect.

NNT is a count-based transformation of a binary event risk difference. It is not a cost-effectiveness ratio, a QALY gain, or the number of treatment-years required; an intervention may prevent events yet still differ greatly in cost, quality of life, and adverse effects. A health economic model should retain the underlying absolute risks, full event pathways, costs, utilities, and uncertainty instead of using NNT as a substitute for those inputs. If treatment effects are estimated observationally, the causal interpretation additionally depends on adequate control of confounding and selection, not merely on computing a risk difference.

Checks before reporting NNT

These checks keep the result interpretable and reproducible. They matter as much as the reciprocal itself because a mathematically correct value can answer the wrong clinical question. State the underlying risks alongside the reported whole-number NNT.

1. Name the event. Define whether the outcome is beneficial or harmful and how it was measured.
2. Name the comparison. Specify treatment, comparator, eligible population, and analysis approach.
3. Fix the horizon. State follow-up time and use time-specific cumulative risks when follow-up or censoring varies.
4. Calculate the absolute difference. Express risks as proportions, verify their direction, and divide one by the nonzero beneficial or harmful difference.
5. Round and label. Round the positive result up and identify it as NNTB or NNTH for the stated event.
6. Present uncertainty. Give the risk-difference interval, handle zero-crossing and infinity correctly, and show relevant baseline-risk variation.
7. Interpret the decision. Consider other benefits, harms, costs, preferences, and the credibility of the effect estimate.

Sources and further study

The current methodological guidance supports the comparative definition, baseline-risk dependence, whole-person rounding, and the discontinuity at zero. The original time-to-event and competing-risk papers extend the calculation to settings where a simple observed proportion is inadequate. The worked figures above are illustrative and have been independently recalculated.

1. Cochrane. Handbook Chapter 15: Interpreting results and drawing conclusions. Authoritative methods guidance for NNTB and NNTH, rounding, baseline risk, and uncertainty.
2. Altman DG. Confidence intervals for the number needed to treat. Original methods discussion of reciprocal confidence limits crossing infinity.
3. Altman DG, Andersen PK. Calculating the number needed to treat for trials where the outcome is time to an event. Original time-specific survival formulation.
4. Number needed to treat for time-to-event data with competing risks. Original extension emphasizing cumulative incidence in the presence of competing events.