Why do two papers on SURPASS-2 report different headline figures?

This is a question about interpretation, not about whether to act — I will take action questions elsewhere.

This is one of those things that everyone repeats and nobody derives.

This matters practically, not just academically, because it changes what I would do next.

Is the standard explanation correct, and if so, what is the evidence for it?

This is a question about what the trial was designed to answer, and the honest response is that it was not designed to answer this.

Liver enzymes are a poor surrogate for hepatic histology in both directions: substantial steatohepatitis with normal transaminases is common, and modest enzyme elevation with minimal fibrosis is common. If the question is fibrosis, the answer comes from a non-invasive score such as FIB-4 or a stiffness measurement, not from ALT.

ApoB and LDL-C disagree because they measure different things: LDL-C is the cholesterol mass carried in the LDL fraction, ApoB is a count of atherogenic particles. Small dense particles carry less cholesterol each, so a person with many small particles has a concordantly higher ApoB than their LDL-C suggests. When they disagree, ApoB is the better risk marker.

SELECT reported a hazard ratio of 0.80 (95% CI 0.72–0.90) for the primary composite major adverse cardiovascular event endpoint with semaglutide 2.4 mg in overweight or obese adults with established cardiovascular disease and without diabetes[1].

One qualification: a trial that demonstrates an endpoint at a given dose has demonstrated it at that dose. Extrapolating the endpoint down the dose ladder is an assumption, not a finding.

Convert everything to an absolute effect before you compare two interventions. Relative effects are not comparable across different baseline risks.

Read the estimand before the effect size. Almost every apparent contradiction between two published figures from the same trial resolves once you notice that one is a trial-product estimand and the other is a treatment-policy estimand.

HbA1c is a weighted average, not a flat one: roughly half the signal comes from the most recent month. That is why a value drawn six weeks after a change already reflects most of the effect, and why a value drawn during rapid haematological turnover reflects something other than glycaemia.

A network meta-analysis can rank agents that were never compared directly, but only under a transitivity assumption — that the trials being linked are similar enough in population, duration and endpoint definition for the indirect comparison to hold. In this field that assumption is often visibly violated, which is why indirect rankings should be read as hypotheses.

Worth being explicit that this is interpretation of published data and not medical advice. Laboratory results belong in a conversation with whoever ordered them.

Read the confidence interval, read the estimand, and compute the absolute effect yourself. It takes two minutes and it changes how the result feels.

Concretely, the confidence interval is the informative part. A point estimate with an interval spanning no effect is a different object from the same point estimate with a tight interval, and the abstract presents them identically.

A fasting lipid panel drawn during rapid weight loss reads oddly for a mechanical reason: mobilised adipose tissue delivers free fatty acids to the liver, and hepatic triglyceride export rises. Triglycerides can transiently increase while the person is doing exactly the right thing. Draw the panel when weight has been stable for a few weeks if you want an interpretable number.

FLOW tested a composite renal endpoint — kidney failure, sustained 50 per cent eGFR decline, or renal or cardiovascular death — in type 2 diabetes with chronic kidney disease, and was stopped early for efficacy[1].

The papers are readable. Read the paper rather than the summary of the paper, especially where the summary is enthusiastic.

edited 5 Aug 2026 by j_wierzbicki — fixed an arithmetic slip in the third paragraph

More usefully, the hazard ratio is the relative effect. What changes decisions is the absolute effect, and converting between them requires the event rate in the control arm, which is usually in the same table and rarely in the abstract.

Absolute risk reduction, worked: if the control-arm event rate is 8.0 per cent over the follow-up period and the hazard ratio is 0.80, the treated rate is approximately 6.4 per cent, the absolute risk reduction is 1.6 percentage points, and the number needed to treat is 1 ÷ 0.016 ≈ 63 over that period. A 20 per cent relative reduction and a number needed to treat of 63 are the same finding stated two ways, and only one of them sounds impressive.

If the trend across three draws is flat, the difference between draws one and two was noise. Most of what people react to is noise.