Number needed to treat calculator, with a confidence interval

Start with the event rate in each arm. The control event rate (CER) is the events in the control arm divided by its participants; the experimental event rate (EER) is the same for the experimental arm. For an unwanted event, the absolute risk reduction (ARR) is CER minus EER. The relative risk (RR) is EER divided by CER, and the relative risk reduction (RRR) is 1 minus RR.

The number needed to treat is the reciprocal of the absolute difference: NNT = 1 / ARR. It reads as the average number of people who need the experimental treatment instead of the control, over the trial's follow-up, for one extra person to avoid the event. By convention it is rounded up, so an ARR of 0.12 gives 8.3, reported as 9, and an ARR of 0.22 gives 4.5, reported as 5 (the Cochrane Handbook uses these examples in chapter 15).

Worked example with the demo numbers: 100 of 500 control participants and 75 of 500 experimental participants have the event. CER = 20.0%, EER = 15.0%, ARR = 5.0 percentage points, RR = 0.75 and RRR = 25%. NNT = 1 / 0.05 = 20. Treat 20 people with the experimental option and, on average, one fewer will have the event than if all 20 had received the control.

An NNT has no meaning without a time frame. An NNT of 20 over one year and an NNT of 20 over five years describe very different treatments, so always state the follow-up period alongside the number. Likewise, an NNT from a high-risk trial population does not transfer to a lower-risk clinic population: apply the trial's relative effect to the new baseline risk instead, as the table above shows.

Use NNTB and NNTH rather than a bare NNT when the direction matters. The Cochrane Handbook recommends these labels and advises against the older "number needed to harm" wording, because readers can misread it as the number of people harmed. For a meta-analysis, do not add up events and participants across trials and divide; compute the NNT from the pooled risk difference, or from a pooled relative effect applied to an assumed control risk.

An NNT is only as good as the event counts behind it. Every event needs a clear definition in the protocol, a form that records it the same way at every site, and a denominator that matches the analysis population. See clinical trial endpoints for how binary and time-to-event endpoints differ, and use the sample size calculator to check that the trial can detect the absolute difference you hope to report.

In Capture, the event forms are eCRFs with edit checks that raise a query automatically when an entry breaks a rule, and every change sits in a field-level audit trail with the reason for change. Allocation can run inside the same system with built-in randomisation, and blinded roles never receive treatment-arm values. At analysis time, CSV or Excel exports come with a data dictionary, and SDTM datasets export as SAS XPT files with Define-XML, so the statistician can calculate the risk difference and its interval in their own validated software. EDC for biostatisticians and SAS programmers covers the export side in more detail.