Accepted answer
Your arithmetic is correct. Your discomfort about the composite is also correct, and it is the more important of the two points.
Checking the numbers
- Semaglutide: 331 / 1767 = 0.1873, so 18.73%.
- Placebo: 410 / 1766 = 0.2321, so 23.21%.
- ARR = 23.21 - 18.73 = 4.48 percentage points.
- NNT = 1 / 0.0448 = 22.3, so 23 treated over the trial horizon to prevent one composite event.
- RRR = 4.48 / 23.21 = 0.193, about 19%, against a reported hazard ratio of 0.76, 95% CI 0.66 to 0.88 [1].
Note that the crude RRR of 19% and the hazard-ratio-implied 24% do not match exactly. That is expected and it is not an error. The hazard ratio is an average instantaneous rate ratio over follow-up from a proportional-hazards model; the crude proportion ratio ignores when events happened and who was censored. When event rates are as high as 20% and follow-up is long, the two diverge noticeably. Neither is wrong; they answer slightly different questions. For an NNT you want the absolute risk difference, so your route is the right one - just do not then quote the hazard-ratio percentage as if it were the same quantity.
On rates per 100 patient-years: those exist because differential follow-up makes crude proportions misleading. In FLOW, follow-up was reasonably balanced, so the crude approach is defensible. As a habit, if the paper reports both proportions and rates and they tell the same story, use proportions for absolute arithmetic. If they disagree, that disagreement is itself information about censoring and you should stop and read the flow diagram.
The composite is the real problem
The FLOW composite bundles onset of kidney failure (sustained eGFR under 15, initiation of dialysis, or transplantation), a sustained 50% or greater reduction in eGFR from baseline, and death from kidney or cardiovascular causes. A patient does not regard a 50% eGFR decline and a cardiovascular death as the same event, so one NNT covering both is an average over things that should not be averaged.
What to do instead: report the composite NNT, then report the components separately and let the reader see the composition. In FLOW the reassuring feature is that the components moved in the same direction rather than one dominating - the kidney-specific components and the death components both favoured semaglutide, and cardiovascular death separately showed a hazard ratio around 0.71, 95% CI 0.56 to 0.89, with all-cause death around 0.80, 95% CI 0.67 to 0.95 [1]. A composite whose components agree is far more interpretable than SELECT's, where stroke went the other way.
The eGFR slope is a different kind of evidence
Yes, and it is worth more attention than it usually gets. FLOW reported a mean annual eGFR slope difference of roughly 1.16 mL/min/1.73 m² per year in favour of semaglutide, 95% CI about 0.86 to 1.47.
Why that is informative independently of the composite: the composite is driven by threshold-crossing events, which are noisy and depend on where the thresholds sit. Slope uses every measurement from every participant and is far more efficient statistically. If a drug slowed decline you would expect both a slope difference and fewer threshold crossings, and seeing both is much stronger than seeing either alone. A composite benefit with a flat slope would suggest the composite was being driven by something other than progressive kidney disease.
The standard caveat applies: GLP-1 receptor agonists cause an early haemodynamic dip in eGFR in some people, so slope analyses in this class are usually reported both from baseline and from a post-initiation timepoint. Check which one you are reading before comparing to another agent's slope.
One structural point
FLOW used semaglutide 1.0 mg, the glycaemic dose, not the 2.4 mg obesity dose. Mean weight loss in FLOW was accordingly modest. Anyone arguing that renal benefit in this class is simply a consequence of large weight loss has to account for a renal composite moving substantially at a dose that does not produce large weight loss.
edited 10 Jun 2025 by sample_id — updated for the 2026 guidance change
The point that a composite with concordant components is more interpretable than one with a discordant component is exactly right. – sian_llewellyn 2 months ago 2The early eGFR dip catches people out when they compare slopes across trials with different reference timepoints. – marcus_thorbjorn 4 months ago add a comment