SELECT reported HR 0.80 for MACE. What does that actually license me to conclude?

I keep seeing SELECT summarised as "semaglutide cuts heart attacks and strokes by 20%". I understand that 20% comes from the hazard ratio of 0.80, and I understand that a hazard ratio is a relative measure, so the 20% is relative to the placebo event rate rather than an absolute 20 percentage points. What I do not have a feel for is what the trial genuinely established versus what is being read into it.

Things I am unsure about, specifically:

• The primary endpoint was a three-part composite. If one component drives the whole result, is it honest to describe the drug as reducing all three?
• The confidence interval was roughly 0.72 to 0.90. Does the fact that it excludes 1.00 tell me the effect is at least 10%, or is that a misreading of what a confidence interval is?
• Several secondary endpoints had intervals that crossed 1.00. I have read those described both as "no effect" and as "underpowered", which cannot both be the default reading.

I would rather understand the inference properly than memorise a number. What claims does this trial support, and what claims are people making that it does not?

Adding the arithmetic, because relative risk is where the intuition goes wrong and absolute numbers fix it.

SELECT event counts, approximately: 569 primary events among 8803 on semaglutide, 701 among 8801 on placebo.

Step by step

• Semaglutide event rate: 569 / 8803 = 0.0646, or 6.46%.
• Placebo event rate: 701 / 8801 = 0.0796, or 7.96%.
• Absolute risk reduction: 7.96 - 6.46 = 1.50 percentage points.
• Relative risk reduction: 1.50 / 7.96 = 0.188, or about 19% - which is where the "20%" comes from, and it agrees with the hazard ratio of 0.80 as it should.
• Number needed to treat: 1 / 0.0150 = 66.7, so 67 people treated for the trial duration to prevent one primary event.

Now attach the time, because an NNT without a horizon is meaningless. Mean follow-up was about 39.8 months, call it 3.3 years. So the statement is: 67 people treated for 3.3 years to prevent one cardiovascular death, non-fatal MI or non-fatal stroke. Equivalently 67 x 3.3 = 221 patient-years of treatment per event prevented.

Two caveats on that arithmetic that matter if you plan to reuse the method:

• Dividing crude counts by randomised totals gives a risk difference, but the hazard ratio came from a time-to-event model that handles censoring. The two agree closely here because follow-up was similar in both arms and event rates are low. In a trial with heavy differential dropout they can diverge, and then you should take the model-based numbers rather than doing this by hand.
• NNT does not scale linearly with time. Extrapolating "67 over 3.3 years" to "20 over 10 years" assumes constant hazard and constant treatment effect, and in SELECT the curves were still diverging at the end of follow-up, so a linear extrapolation could be conservative or optimistic depending on whether that divergence continues.

For calibration against familiar therapy: statins in secondary prevention sit in a broadly similar NNT range over comparable horizons. That is the right comparison to have in mind - a real, worthwhile, unspectacular population-level effect, not a transformation of individual risk.

On the weight-mediation question, since it is the part of SELECT people argue about most and the trial gives partial answers.

Arguments that the benefit is not purely weight-mediated:

• Curve separation began within the first few months, before much of the eventual mean weight loss of roughly 9 to 10% had accumulated [1].
• The benefit was broadly consistent across baseline BMI strata, including the lowest stratum, where absolute weight loss was smallest.
• Prespecified analyses reported benefit across categories of achieved weight change, including in participants who lost relatively little [2].
• Effects on inflammatory markers, blood pressure and lipids appear earlier and larger than weight change alone would predict.

Arguments for caution before concluding a direct vascular effect:

• Subgroup consistency across achieved weight loss is a post-randomisation stratification. Achieved weight loss is not randomised, so comparing across those categories is observational and open to confounding by adherence and by whatever makes a person a poor responder.
• Early separation is compatible with early haemodynamic and metabolic changes that are themselves consequences of reduced intake, not only of fat mass.
• Formal mediation analysis in this setting depends on modelling assumptions that are not testable within the trial.

The parallel evidence from SUSTAIN-6, where semaglutide at glycaemic doses of 0.5 and 1.0 mg reduced MACE with a hazard ratio of 0.74, 95% CI 0.58 to 0.95, at weight losses far smaller than SELECT's, is probably the strongest single argument that weight is not the whole story [3]. It is also worth remembering SUSTAIN-6 was designed as a non-inferiority safety trial, so its superiority finding is a bonus rather than a confirmatory result, and its retinopathy-complication signal with a hazard ratio near 1.76, 95% CI 1.11 to 2.78, is part of the same dataset and should not be quietly dropped when the trial is cited approvingly.