How to Read a GLP-1 Study

This is a slightly different piece to the rest of the site, and I think it is the most durable thing here. Individual findings go out of date. Knowing how to read them does not.

Everything below can be done in about the time it takes to read the press release that sent you looking.

1. Find the absolute numbers

This is the highest-yield check and it takes thirty seconds.

Relative risk reduction is the proportional difference between groups. Absolute risk reduction is the difference in actual event rates. They describe the same finding and they feel completely different.

If a trial reports events falling from 10% to 8%, that is:

• a 20% relative risk reduction
• a 2 percentage point absolute risk reduction
• a number needed to treat of 50 — fifty people treated for one to benefit

All three are true. Headlines use the first, because it is the biggest number. The second and third are the ones that tell you what it means for a person.

Neither framing is dishonest, and using only the relative figure is misleading by omission. A 20% relative reduction on a common outcome is a major public health finding. The same 20% on something that happens to 1 person in 10,000 is nearly nothing.

So: whenever you see a percentage reduction, ask “of what?” The absolute rates are almost always in the paper even when they are absent from the coverage.

2. Ask who was in the trial

Inclusion criteria are usually in the abstract and always in the methods, and they determine who the result actually describes.

Take SELECT: over 17,000 participants, aged 45 and over, with overweight or obesity and established cardiovascular disease, without diabetes. That is a specific, higher-risk population. The finding describes them well.

Now consider a healthy 34-year-old with a BMI of 31 and no cardiovascular disease. Is the finding relevant? Probably in direction. Almost certainly not in size — and the absolute benefit will be smaller, because their baseline risk is smaller. The same relative reduction applied to a lower starting risk yields less.

Check age, BMI range, diabetes status, and what was excluded. Trials routinely exclude people with the conditions that make treatment complicated, which is exactly why real-world results tend to be less tidy.

3. Ask what was measured

Hard endpoints are things that matter in themselves: death, heart attack, stroke, kidney failure, hospitalisation, fracture.

Surrogate endpoints are stand-ins: HbA1c, LDL cholesterol, blood pressure, a change in body weight, a score on a scale.

Surrogates are not worthless — they are faster and cheaper, and often the only practical option. They are also not guaranteed to translate, and medicine has a long history of drugs that moved a surrogate impressively and did nothing, or harm, to the outcome it stood for.

In this field the useful implication is specific: weight change is a surrogate. A trial showing 15% weight loss is telling you about weight. Trials like SELECT and FLOW, reporting cardiovascular events and kidney failure, are telling you about outcomes — which is why they carry more weight than their headline numbers alone would suggest, as discussed in what we know about long-term safety.

Watch also for composite endpoints, where several outcomes are bundled. Check what is inside: a composite driven mainly by its least serious component is a weaker result than the headline implies.

4. Check what kind of document you are reading

In descending order of reliability:

The last row matters in this field specifically. The thyroid boxed warning derives from rodent studies, and the entire question is whether it translates — a question rodent data cannot answer. That is covered in the thyroid cancer warning explained.

Observational studies deserve a note too. They cannot establish causation, and in this area they carry a particular confounder: people on these drugs see clinicians more often and get more tests, so more things get found. Looking harder finds more.

5. Look for the pre-specified primary endpoint

Good trials declare in advance what they are measuring, in a public registry, before enrolling anyone.

What you want to see is that the result being reported is the one that was pre-specified. When a paper leads with a secondary analysis, a subgroup, or an outcome that does not match the registration, that is worth noticing — not necessarily wrong, and a much weaker basis for a claim.

Subgroup findings in particular are hypothesis-generating rather than conclusive. If a trial reports no overall effect but a striking benefit in one subgroup, that subgroup needs its own trial before anyone acts on it.

6. Read the limitations section

It is usually near the end of the discussion, it is written by the authors, and it is frequently the most honest paragraph in the paper. Authors know their study’s weaknesses better than any critic.

Also worth checking: who dropped out and why. Discontinuation rates in this field are substantial, and how a trial handled missing data can move the headline number. A result presented for people who completed the trial describes a different population to one that includes everyone randomised.

The industry funding question

Nearly every large trial in this field is manufacturer-funded, because a 17,000-person outcome trial costs more than any academic funder will provide. A rule of ignoring industry-funded research would leave you with almost no evidence.

So the question is not whether it was funded, but whether it was done in a way that constrains the funder: pre-registration, an independent data monitoring committee, blinded adjudication of outcomes, peer review, and full publication of the protocol. Those are the structural protections, and they are checkable.

Be more sceptical of the framing than the data. The trial is usually sound; the press release is often not.

A worked example

Suppose you read: “New study shows GLP-1 drug cuts heart attacks by 20%.”

• Absolute numbers? Not in the headline. In the paper, events fell from 8.0% to 6.5% — a 1.5 percentage point absolute reduction over about 3 years.
• Who? People with established cardiovascular disease. Higher baseline risk than the general population.
• What was measured? A composite of cardiovascular death, non-fatal heart attack and non-fatal stroke. Hard endpoints, which is good — check which component drove it.
• What kind of document? Peer-reviewed, published, pre-registered.
• Pre-specified? Yes, the primary endpoint.
• Limitations? Specific population, about 3 years of follow-up, manufacturer-funded with independent adjudication.

Conclusion: a genuine, meaningful result in a defined population, smaller in absolute terms than the headline implies, and less directly applicable the further you are from the trial population.

That took two minutes, and it is a more accurate understanding than most of the coverage will give you. The same six checks work on the next study too — including on the ones I cite in the trial table, which you are welcome to apply them to.