Number Needed to Harm, and How to Read It Against Number Needed to Treat

Number needed to harm is the count of people who would have to take a treatment before one of them has a particular side effect caused by it. If a medicine produces one extra case of a problem for every 200 people who take it, its number needed to harm for that problem is 200. On its own the figure is only half a story. Its partner, number needed to treat, counts how many people take the same medicine before one gains the benefit being measured. Set the two counts next to each other and a fuzzy "helps some, hurts some" resolves into two plain numbers you can actually weigh. The weighing itself belongs to you and a clinician who knows your situation.

Start with the benefit, then demand the harm#

Most treatment claims arrive benefit-first, and the benefit is usually the number people quote from memory. Number needed to treat captures it: one divided by the absolute risk reduction, the extra chance of avoiding a bad outcome because of the treatment. A number needed to treat of 25 means that across 25 people treated, one avoids the tracked outcome. The other 24 take the treatment without that visible gain, which is normal and expected, not a failure.

The habit worth building is to refuse to stop there. For every benefit count, ask for the matching harm count before forming a view. Number needed to harm supplies it by the same arithmetic in the opposite direction: one divided by the absolute risk increase, the extra chance of a side effect caused by the treatment. Because it reads backward from intuition, it pays to say it out loud: a number needed to harm of 20 is worse than one of 2,000, since 20 means harm arrived after only a handful of people were treated.

Both figures are the inverse of an absolute difference, so both inherit everything that difference depends on. Change the baseline risk of the group, the length of follow-up, or the exact outcome counted, and the number moves. That is why a bare "the number needed to harm is 200" tells you nothing until you also know which harm, compared with what, over how long.

A worked example with invented numbers#

To keep anything real from riding on the illustration, here are round invented figures. Picture a medicine tested in 800 people against a comparison group of 800, followed for two years.

On the benefit ledger, the outcome the medicine is meant to prevent occurs in 60 of the comparison group but only 36 of the treated group. That is an absolute risk reduction of 24 in 800, or 3 percent, which gives a number needed to treat of about 33.

On the harm ledger, a bothersome side effect appears in 40 of the treated people and 16 of the comparison group. That is an absolute risk increase of 24 in 800, again 3 percent, so the number needed to harm is also about 33.

When the two counts land near each other like this, the trade-off is stark: for roughly every person spared the outcome, one acquires the side effect. Whether that trade is worth making has nothing to do with the arithmetic and everything to do with the two outcomes. One avoided stroke set against one mild, passing rash is an easy yes. One avoided rash set against one serious bleed is an easy no. The counts pose the question cleanly; they do not answer it.

Why the two numbers have to travel together#

A benefit shown without its harm invites a lopsided decision, and a harm shown without its benefit invites needless alarm. An honest summary states both counts in one breath, over one timeframe, in one population. A benefit measured across five years cannot be laid against a harm measured across six months, and both should come from comparable groups, ideally the same trial. Otherwise the ledger is tilted before you read a single figure.

Weighting then does the rest of the work. A number needed to harm of 33 for something fleeting and reversible is not the same as the identical figure for something lasting and severe. Reasonable people value outcomes differently, which is exactly why this calculation ends in a conversation rather than a verdict. The arithmetic narrows the choice; your priorities close it.

Read any benefit-and-harm claim in under a minute#

You need no statistics training for this, only a fixed order of questions. First, ask for the benefit as an absolute count: out of 1,000 people treated for a stated time, how many gain? Then ask the same of the harm, and how bad the harm is. Watch for mixed framing, where the upside is dressed up as a big relative number ("cuts the risk in half") while the downside is played down as a small absolute one ("rarely causes a problem"). Those two are measured on scales that do not compare. Convert both to counts out of the same denominator and the flattering asymmetry disappears.

One caution keeps these figures from being oversold. Many harms only show up once a treatment is used widely and for longer, so an early number needed to harm can look reassuring merely because the harm has not been seen yet. "None observed" is not "none exists." Larger populations and more time routinely revise these numbers upward.

Where the trade-off bites hardest#

Two settings deserve extra care. The first is prevention in people who currently feel well: the benefit is often a small absolute reduction spread thinly across many people, which pushes the number needed to treat high and makes even a modest harm loom larger by comparison. The second is any treatment taken for years, where small yearly harms steadily accumulate while the benefit may level off.

None of this makes side effects a reason to reject a genuinely useful treatment, nor a measured benefit a reason to wave away real harm. The single recurring error is reading one number without its partner. So keep the routine simple. For any treatment, ask how many people it helps and how many it harms, over the same time, in people like you, and how much the help and the harm each weigh. A recommendation that answers all four earns your trust. One that offers only the cheerful half is unfinished.