Why do two papers on SURMOUNT-4 report different headline figures?

I have three data points across nine months, which I hope is enough to see a trend.

I understand the observation; what I do not understand is the mechanism behind it.

I have read the two review articles that come up first and both assert this without a citation to a primary source.

So what is the mechanism, and how well established is it?

Because two papers on SURMOUNT-4 are usually reporting two different estimands from the same randomisation. The treatment-policy estimand asks what happened to everyone assigned, including those who stopped; the trial-product estimand asks what happens if you keep taking it. The second is always the larger number, and both are legitimate answers to different questions. Then there is the analysis population — randomised, treated, or completers — and the handling of missing data, where a last-observation-carried-forward and a multiple imputation can differ by a point or more. Neither paper is wrong. Read the statistical methods section and you will find both figures defined in it.

Start with what the trial was powered for. Everything else in the publication is secondary, exploratory, or a subgroup, and those three words mean three different things.

Trial populations are selected. Exclusion criteria in this class routinely remove people with significant renal impairment, prior pancreatitis and unstable psychiatric illness, which is exactly the population the results are then quoted for.

Open-label extensions are not the same evidence as the randomised phase. Once everyone knows what they are taking, the reported outcomes acquire a bias that no analysis fully removes.

The caveat is that trial evidence is about licensed product administered under supervision. None of it transfers automatically to research-grade material of unverified content.

When two sources disagree, the answer is almost always in the methods section of the one you have not read.

Before comparing two trials, check whether they share an endpoint definition. Frequently they do not, and the numbers then are not comparable in any sense.

Duration decides what can be seen. A 68-week trial can measure weight and glycaemia; it cannot measure anything whose event rate is one per cent per year without enrolling tens of thousands.

Non-inferiority and superiority designs are not interchangeable. A non-inferiority result says the new agent is not meaningfully worse against a pre-specified margin — it does not say it is as good, and it certainly does not say it is better.

I am not a clinician and this is not medical advice; it is a reading of a published protocol.

Quote the interval alongside the estimate and half the disagreements on this site would not start.

This is answerable from the published record, but only if you take the placebo arm seriously rather than reading the active arm alone.

Confidence intervals matter more than point estimates when two trials disagree. Two studies reporting fifteen and twenty per cent whose intervals overlap heavily have not disagreed about anything.

Placebo arms in this class are not nothing. Lifestyle-intervention placebo arms in the major obesity trials commonly lose two to three per cent of body weight, so an active-arm figure quoted without its comparator overstates the drug effect by roughly that much.

Be careful about generalising from a trial population to yourself. The exclusion criteria are usually the most informative page in the supplement.

Read the protocol and the statistical analysis plan if the result matters to you. Both are usually published alongside.

The part that matters: a trial establishes what happened to a defined group under a defined protocol. Extending it beyond that group is inference, and inference is allowed as long as it is labelled.

Intention-to-treat and per-protocol analyses answer different questions. ITT asks what happens if you offer the treatment; per-protocol asks what happens if it is taken as directed. The gap between the two is a measure of how tolerable the protocol was.

Registry entries at ClinicalTrials.gov carry the pre-specified primary endpoint with a timestamp, which is the cheapest available check on whether an endpoint was changed after the data were seen.

The short version: check the endpoint, check the comparator, check who was excluded, then look at the number.

Answer first: read the primary endpoint, the comparator and the population before you read the effect size. Almost every argument on this site about a trial is really an argument about one of those three.

A composite endpoint is only as informative as its least serious component. Where a cardiovascular composite combines death, infarction and stroke, ask which component moved, because they are not interchangeable outcomes.

Meta-analyses in this area are dominated by whichever trial contributed the most participants, so read the forest plot rather than the summary estimate.

If a claim cannot be traced to a named trial with a named endpoint, treat it as a claim rather than as evidence.