How Many People to Treat to Help One

Trials almost always report benefit as a percentage reduction, which sounds impressive and tells you nothing about your own odds. The FLOW cardiovascular analysis did something rarer: it published the number needed to treat — how many people take the drug for three years so that one of them avoids the outcome.^[1] The answers were 22, 13 and 17.

What number needed to treat means

A 26% reduction in risk is a ratio. It describes how much the odds shift, not how likely the event was to begin with. Cutting a 40% risk by a quarter is a very different proposition from cutting a 0.4% risk by a quarter, and the same percentage describes both.

Number needed to treat collapses that into one figure: treat this many people for this long, and one of them will avoid the event who otherwise would not have. It is the statistic that answers the question a patient is actually asking, and it is the one most papers leave out.

The trial and the subgroups

Everyone in FLOW carried two diagnoses at once — type 2 diabetes, plus kidney disease already established — and half of them drew weekly semaglutide at 1.0 mg. Its primary outcome was a composite of serious kidney deterioration — a halving of filtration rate, filtration falling below 15, dialysis, transplantation — together with kidney or cardiovascular death. This analysis split the trial by cardiovascular status at entry.

All-cause death moved in the same direction throughout, with hazard ratios between 0.71 and 0.82 across every subgroup and no sign of heterogeneity.

Why some intervals cross 1 and it does not matter here

Read down the hazard ratio columns and several intervals include 1 — atherosclerotic disease at 0.63 to 1.02, all-cause death in the heart-failure group at 0.54 to 1.05. Taken alone, each of those says “not established.”

That is exactly what splitting a trial does. Halve the participants and you halve the events, which widens every interval regardless of whether the underlying effect changed. Subgroup intervals crossing 1 in a trial whose overall result was positive is the expected picture, not a warning sign.

That is the mirror image of a case we covered separately, where an interaction P of .06 sat under a table that looked like a clean split — the statistic that asks if groups differ. Same test, opposite reading, and the pair is more instructive than either.

Why you cannot carry these numbers elsewhere

So these figures say something precise and narrow: for people with diabetes and kidney disease, treating between 13 and 22 for three years prevents one serious kidney outcome. Applied to someone taking a GLP-1 for weight without kidney disease, they would be fabricated. We are not going to extrapolate them and nobody else should either.

The dose is worth noting too: semaglutide 1.0 mg, the type 2 diabetes strength, not the 2.4 mg used for weight management. For the kidney evidence more broadly, see GLP-1s and your kidneys.