Number Needed to Treat

The number needed to treat (NNT) is a foundational clinical impact metric in evidence-based medicine, clinical epidemiology, biostatistics, and public health that translates abstract probability differences from randomized controlled trials into an intuitive integer scale: the average number of patients who must be treated with a specific intervention (rather than a control or comparator) over a defined follow-up duration to prevent one additional bad clinical outcome (or, for therapeutic cure endpoints, to achieve one additional favorable outcome). Introduced in 1988 by Canadian clinical epidemiologists Andreas Laupacis, David L. Sackett, and Robin S. Roberts in the New England Journal of Medicine, the NNT is defined mathematically as the reciprocal of the Absolute Risk Reduction (ARR). By anchoring therapeutic benefit to absolute baseline risk rather than purely relative percentage reductions, the NNT—together with its companion metrics, the Number Needed to Harm (NNH) and the Number Needed to Vaccinate (NNV)—prevents the widespread cognitive distortion caused by reporting Relative Risk Reduction (RRR) in isolation and enables direct bedside and policy weighing of clinical benefits, drug toxicities, and healthcare resource allocation.

Derivation from Two-Arm Clinical Trial Incidence Proportions

Consider a randomized controlled trial or prospective cohort study followed over a fixed time horizon T (for example, T = 5 years), comparing an active Experimental Treatment Group (E) against an untreated, placebo, or standard-of-care Control Group (C):

• Let I_C (also written CER, Control Event Rate) denote the cumulative incidence proportion (attack rate or baseline risk) of the adverse outcome (such as myocardial infarction, stroke, or death) in the control group over period T.
• Let I_E (also written EER, Experimental Event Rate) denote the cumulative incidence proportion of the adverse outcome in the treated group over the same period T.

From these two event probabilities (I_C and I_E), epidemiologists compute four distinct effect-size measures:

1. Relative Risk (Risk Ratio, RR): The ratio of the risk in the treated group to the risk in the control group:

RR = I_E / I_C

1. Relative Risk Reduction (RRR): The proportional percentage by which the treatment reduces risk relative to baseline:

RRR = ( I_C − I_E ) / I_C = 1 − RR

1. Absolute Risk Reduction (ARR, or Risk Difference, RD): The simple arithmetic difference between the control event rate and the experimental event rate:

ARR = I_C − I_E = I_C × RRR

1. Number Needed to Treat (NNT): The multiplicative inverse (reciprocal) of the Absolute Risk Reduction:

NNT = 1 / ARR = 1 / ( I_C − I_E ) = 1 / ( I_C × RRR )

In clinical reporting, fractional NNT values are conventionally rounded up to the next highest whole integer (for example, NNT = 14.2 is reported as NNT = 15), because a fraction of a patient cannot be treated and rounding down would slightly overstate clinical efficacy. An ideal, universally curative treatment for an otherwise 100% fatal condition (I_C = 1.0, I_E = 0.0, ARR = 1.0) achieves the theoretical minimum of NNT = 1 (every single treated patient is saved). As ARR approaches zero, NNT approaches infinity (∞), indicating zero net benefit.

Why Identical Relative Risk Reductions Yield Vastly Different NNTs (The Baseline Risk Effect)

The decisive epidemiological insight of the NNT is unveiled by the identity NNT = 1 / (I_C × RRR): while the pharmacological Relative Risk Reduction (RRR) of a drug (such as an HMG-CoA reductase inhibitor statin or an antihypertensive agent from the WHO list of essential medicines) often remains roughly constant (~25% to 30%) across diverse patient populations, the Absolute Risk Reduction (ARR) and NNT vary by orders of magnitude depending on the patient's baseline untreated risk (I_C):

As this comparison demonstrates, quoting only that a drug "reduces heart attacks by 33.3% (RRR)" conceals a 500-fold difference in clinical efficiency between Cohort A (NNT = 10) and Cohort D (NNT = 5,000).

1. Number Needed to Harm (NNH) and Likelihood of Being Helped or Harmed (LHH)

Virtually every pharmacological or surgical intervention carries a non-zero risk of adverse drug reactions or procedural complications (such as major gastrointestinal bleeding from aspirin/anticoagulants, myopathy or new-onset diabetes from high-intensity statins, or immune-related colitis from checkpoint inhibitors). If the incidence of an adverse side effect is I_harm,E in the treated group and I_harm,C in the control group, the Absolute Risk Increase (ARI) is ARI = I_harm,E − I_harm,C, and its reciprocal is the Number Needed to Harm (NNH):

NNH = 1 / ARI = 1 / ( I_harm,E − I_harm,C )

Whereas clinicians desire a small NNT (ideally 1 to 50), they desire a large NNH (ideally > 200 to 10,000, meaning hundreds or thousands of patients can be treated before one additional patient suffers a treatment-induced adverse event).

To combine benefit and harm into a single bedside trade-off ratio, Straus and Sackett defined the Likelihood of Being Helped vs. Harmed (LHH):

LHH = ( 1 / NNT ) : ( 1 / NNH ) = ARR : ARI = NNH / NNT

*(Weighted optionally by the patient's relative utility disvalue s comparing the severity of the disease event against the severity of the drug side effect.)* For example, in secondary prevention of myocardial infarction with low-dose daily aspirin (NNT ≈ 50 over 2 years to prevent 1 vascular event; NNH ≈ 250 to cause 1 major extracranial GI bleed), LHH = 250 / 50 = 5 : 1 in favor of treatment. Conversely, in low-risk primary prevention where baseline vascular risk is minimal (NNT ≈ 1,200; NNH ≈ 250), LHH drops to 0.21 : 1 (the patient is nearly 5 times more likely to suffer a major bleed than to avert a heart attack), explaining why modern guidelines removed routine daily aspirin from low-risk primary prevention.

2. Number Needed to Vaccinate (NNV) and Indirect Herd Immunity Amplification

In immunization policy, the NNT framework is adapted as the Number Needed to Vaccinate (NNV)—the average number of individuals who must receive a vaccine to prevent one clinical case, one hospitalization, or one death over a defined period:

NNV_direct = 1 / ( AR_unvaccinated × VE )

However, as emphasized by Tuite and Fisman, applying the simple static formula NNV_direct to contagious diseases understates the true population benefit of vaccination whenever coverage approaches the herd immunity threshold. Because each vaccinated person who avoids infection also breaks chains of onward transmission to unvaccinated contacts (indirect / herd protection), dynamic transmission models show that the effective population NNV is substantially smaller (more favorable) than NNV_direct during active outbreaks.

1. Time-Horizon Dependence (Why Every NNT Must Specify T)

An NNT is mathematically uninterpretable unless accompanied by its explicit time horizon T. Chronic preventive therapies (such as antihypertensives, statins, or bisphosphonates) accumulate prevented events continuously over time: if a treatment reduces an annual event hazard of 2% per year by 25% (ARR_1-yr = 0.5% per year), its 1-year NNT is 200, its 5-year NNT is ~43, and its 10-year NNT is ~23. To compare trials of unequal follow-up length (T₁ vs. T₂), analysts standardize to a common duration under constant hazard assumptions or compute person-years of treatment needed to prevent one event.

2. Douglas Altman's Confidence Interval Across Infinity (NNTB to ∞ to NNTH)

As demonstrated by biostatistician Douglas G. Altman (1998) in the BMJ, the NNT exhibits unusual statistical geometry when a clinical trial result is not statistically significant (p > 0.05)—meaning the 95% confidence interval for the Absolute Risk Reduction (ARR) crosses zero (spanning from a negative lower bound −L, representing net harm, to a positive upper bound +U, representing net benefit).

Because the reciprocal function f(x) = 1/x has a vertical asymptote (discontinuity) at x = 0, the 95% confidence interval of the NNT does not consist of the numbers between −1/L and +1/U; rather, it consists of all values outside that central interval, passing continuously through infinity (∞)! To avoid confusion with negative numbers, Altman established the standard notation separating NNTB (Number Needed to Treat for Benefit) from NNTH (Number Needed to Treat for Harm):

• Suppose a trial yields ARR = +10.0% (NNTB = 10) with a non-significant 95% CI for ARR of −5.0% to +25.0%.
• At the upper bound (ARR = +25.0%), the best-case benefit is NNTB = 4.
• At ARR = 0.0%, the effect is NNT = ∞.
• At the lower bound (ARR = −5.0%), the worst-case harm is NNTH = 20.
• The correct 95% confidence interval is written: NNTB = 10 (95% CI: NNTB 4 to ∞ to NNTH 20)—never "(−20 to +4)", because values such as NNT = 1, 2, or 3 represent stronger treatment effects that lie outside the confidence interval, whereas values approaching ∞ represent the null hypothesis (ARR = 0) included inside the confidence interval.

3. Calculating NNT from Kaplan-Meier Survival Curves and Cox Hazard Ratios

When time-to-event clinical trials suffer from variable follow-up and right-censoring, computing NNT from crude event counts (a/N₁ − c/N₀) is biased. Instead, as shown by Altman and Per Kragh Andersen (1999), the NNT at landmark time t must be calculated directly from the Kaplan-Meier survival probabilities S_E(t) and S_C(t):

NNT(t) = 1 / [ S_E(t) − S_C(t) ]

Or, when a Cox proportional hazards model reports an adjusted Hazard Ratio (HR) and the baseline survival probability in the control group at time t is S_C(t), the survival probability in the treated group is S_E(t) = [S_C(t)]^HR, yielding the exact proportional-hazards NNT formula:

NNT(t) = 1 / ( [ S_C(t) ]^HR − S_C(t) )

1. Laupacis, A., Sackett, D. L., & Roberts, R. S. (1988). "An assessment of clinically useful measures of the consequences of treatment." New England Journal of Medicine, 318(26), 1728–1733. https://doi.org/10.1056/NEJM198806303182605
2. Cook, R. J., & Sackett, D. L. (1995). "The number needed to treat: a clinically useful measure of treatment effect." BMJ, 310(6977), 452–454. https://doi.org/10.1136/bmj.310.6977.452
3. Altman, D. G. (1998). "Confidence intervals for the number needed to treat." BMJ, 317(7168), 1309–1312. https://doi.org/10.1136/bmj.317.7168.1309
4. Altman, D. G., & Andersen, P. K. (1999). "Calculating the number needed to treat for trials where the outcome is time to an event." BMJ, 319(7223), 1492–1495. https://doi.org/10.1136/bmj.319.7223.1492
5. Mendes, D., Alves, C., & Batel-Marques, F. (2017). "Number needed to treat (NNT) in clinical literature: an appraisal." BMC Medicine, 15(1), 112. https://doi.org/10.1186/s12916-017-0875-8