Is a wide interval the researchers' fault?

Usually not. It reflects how much information the available studies contained. Reporting it honestly, as this paper did, is the correct behaviour. The failure happens downstream, when the point estimate is quoted without it.

Does I-squared of 98% mean the studies are wrong?

No, it means they differ from each other far more than chance explains - because of different populations, comparators, outcome definitions or designs. Any of those can be individually reasonable while making the studies unpoolable.

Why does randomisation help so much here?

Because the alternative explanation for any association between these drugs and psychiatric outcomes is that people who receive them differ from people who do not, in ways related to the outcome. Randomisation removes that by construction; statistical adjustment in observational data can only remove what was measured.

Should observational studies be ignored?

No - they answer questions trials cannot, including what happens over longer periods and in people trials exclude. The point is to match the strength of a claim to the design that produced it, and a four-study pool with 98% heterogeneity supports a very weak claim.

GLP-1 & Incretin Science

A Risk Ratio of 0.57 That Means Almost Nothing

JMWritten & reviewed by Jack Muncaster · Founder, UK PeptidesLast reviewed 2026-08-242 cited sources

A point estimate quoted without its interval is close to meaningless. A 2025 meta-analysis reported a risk ratio of 0.568, which looks protective, with a confidence interval spanning a 92% reduction to a 320% increase, and heterogeneity of 98% between its studies.

Key facts

Reported risk ratio: 0.568
95% confidence interval: 0.077 to 4.205
Lower bound in plain terms: 92% lower risk
Upper bound in plain terms: 320% higher risk
Heterogeneity: I-squared = 98%
Studies pooled in that estimate: 4
Studies screened: 126
Source: Bushi G et al., Diabetes Metab Res Rev 2025 (PMID 39945396)

The number and its interval

Bushi and colleagues screened 126 studies, included 11, and pooled four of them for a meta-analysis of suicidal outcomes among GLP-1 receptor agonist users against users of other anti-hyperglycaemic drugs. The result was a risk ratio of 0.568, with a 95% confidence interval of 0.077 to 4.205. Quoted alone, 0.568 reads as a 43% reduction in risk and would make a reassuring headline. Quoted with its interval, it says the data are compatible with the drugs reducing risk by 92%, with them increasing it more than fourfold, and with everything in between including no effect at all.

What a confidence interval is actually reporting

It is a statement about the precision of an estimate, not about the probability that a particular value is true - a distinction that is routinely blurred and worth keeping. The practical reading is this: values inside the interval are those the data cannot reasonably exclude. An interval that contains 1 means no effect cannot be excluded. An interval running from 0.077 to 4.205 additionally means that a large protective effect and a large harmful effect cannot be excluded either. The width of the interval is the finding here, far more than its centre.

Why the point estimate misleads more than a null result would

A study reporting no difference invites the correct conclusion: this did not settle the question. A study reporting 0.568 invites a wrong one, because a specific number carries an air of measurement. The centre of a very wide interval is not a best guess in any useful sense - it is where the arithmetic happened to land given a handful of studies that disagreed with each other. The asymmetry in how these get reported is predictable: 0.568 travels, 0.077-to-4.205 does not.

The heterogeneity statistic is the second warning

I-squared estimates what share of the variation between studies exceeds what chance alone would produce. Values above 75% are conventionally described as considerable heterogeneity. This analysis reported 98%. At that level the included studies are not measuring the same thing in the same way, and pooling them produces an average of incommensurable quantities. A high I-squared does not merely widen the interval - it undermines the premise that a single summary number exists to be estimated. The authors say so, noting that the high heterogeneity and reliance on pharmacovigilance data suggest caution.

What did settle the question, and why

The FDA's review of the same question used a different kind of evidence: a meta-analysis of 91 placebo-controlled clinical trials across GLP-1 receptor agonist development programmes, comprising 107,910 participants of whom 60,338 received a GLP-1 receptor agonist and 47,572 received placebo. Randomised allocation removes confounding by indication; a hundred thousand participants supplies precision; and pooling trials with comparable designs keeps heterogeneity manageable. Four heterogeneous observational studies could not answer this. Ninety-one randomised trials could. The difference is design and scale, not the direction of the result.

A habit worth forming

Whenever a relative risk, odds ratio, hazard ratio or risk ratio is quoted, ask for the interval before doing anything else with the number. If the interval is not given, the number should not be used. If the interval crosses 1, the study did not demonstrate a difference regardless of where the point estimate sits. And if the interval spans an order of magnitude, the honest summary is that the analysis was uninformative - which is a legitimate finding and a much better one to report than a spurious point estimate.

Quick reference

	Bushi 2025 meta-analysis	FDA trial-level meta-analysis
Evidence type	Observational studies	Placebo-controlled randomised trials
Studies pooled	4 (of 11 included, 126 screened)	91
Participants	Not pooled as a single figure	107,910
Treated / comparator	Mixed comparators	60,338 / 47,572
Heterogeneity	I-squared 98%	Comparable trial designs
Conclusion supported	Uninformative	No increased risk detected

Extended research context

The GLP-1 & Incretin Science deep dive

Deep dive: the two routes to a bigger effect

Every compound trying to beat GLP-1 alone has taken one of two routes. The first adds more receptors from the same hormone family — GIP in tirzepatide, GIP and glucagon in retatrutide. The second adds a non-incretin satiety hormone, which in practice means amylin: CagriSema combines cagrilintide with semaglutide, and amycretin engages both receptors from one molecule. Both routes work, because they recruit signalling pathways that do not fully overlap. Neither has escaped the constraint that binds all of them, which is that gastrointestinal tolerability worsens as effect size grows.

Deep dive: why a percentage is not a result

The most-quoted numbers in this field are the least comparable. REDEFINE 1 reported 22.7% and 20.4% for the same compound in the same trial — the first among participants who adhered to treatment, the second across everyone randomised. TRIUMPH-1 reported 28.3% in an uncomplicated obesity population while TRIUMPH-3 reported up to 22.6% in adults with established cardiovascular disease, using the same compound. Before any two figures can be compared they have to match on estimand, population, duration, comparator and whether the number is placebo-adjusted. Most published comparisons match on none of them.

Deep dive: what happens after the trial stops

Every headline figure describes weight while treatment continues. The STEP-1 extension found that a year after semaglutide was stopped, participants had given back roughly two-thirds of what they lost, moving from 17.3% mean reduction to a net 5.6% — though average weight remained below baseline and nearly half stayed at least 5% down. Meta-analysis puts regain at around 0.8 kg per month. This is why maintenance studies such as TRIUMPH-6 matter more to the field's future than another two points of peak reduction.

Research applications

▸Comparing incretin and amylin compounds on a like-for-like basis
▸Interpreting estimands, thresholds and placebo-adjusted figures in trial reports
▸Tracking the obesity pipeline across sponsors and jurisdictions
▸Understanding receptor pharmacology behind GLP-1, GIP, glucagon and amylin
▸Distinguishing licensed medicines from investigational compounds

Handling checklist

✓Identify which estimand a quoted percentage comes from before citing it
✓Check the trial population and baseline BMI against the comparison you are making
✓Confirm the duration and whether the reduction curve had plateaued
✓Read discontinuation rates alongside efficacy figures
✓Verify every NCT identifier against ClinicalTrials.gov rather than secondary reporting

Common research-handling mistakes

Learnt from thousands of researcher orders across our UK labs.

✗ Comparing headline percentages across different trials

Fix: Population, duration, estimand and comparator all differ; the numbers are not interchangeable.

✗ Quoting the larger of two figures from the same trial

Fix: Name the estimand. Efficacy and treatment-policy answer different questions.

✗ Treating peak reduction as a durable outcome

Fix: Substantial regain follows cessation across the class; peak figures describe a maintained state.

✗ Assuming an oral route means a weaker mechanism

Fix: Route and receptor count are independent. Orforglipron is weaker because it hits one receptor, not because it is a tablet.

✗ Reading investigational compounds as available treatments

Fix: Most of this pipeline holds no authorisation anywhere; mazdutide is approved only in China.

Continue researching

Peer-reviewed guides, comparators and matched reference materials.

Frequently asked questions

Is a wide interval the researchers' fault?: Usually not. It reflects how much information the available studies contained. Reporting it honestly, as this paper did, is the correct behaviour. The failure happens downstream, when the point estimate is quoted without it.
Does I-squared of 98% mean the studies are wrong?: No, it means they differ from each other far more than chance explains - because of different populations, comparators, outcome definitions or designs. Any of those can be individually reasonable while making the studies unpoolable.
Why does randomisation help so much here?: Because the alternative explanation for any association between these drugs and psychiatric outcomes is that people who receive them differ from people who do not, in ways related to the outcome. Randomisation removes that by construction; statistical adjustment in observational data can only remove what was measured.
Should observational studies be ignored?: No - they answer questions trials cannot, including what happens over longer periods and in people trials exclude. The point is to match the strength of a claim to the design that produced it, and a four-study pool with 98% heterogeneity supports a very weak claim.

Primary sources & clinical trials

Peer-reviewed research and registered trials from PubMed, ClinicalTrials.gov, PubChem, FDA and NIH. All links open in a new tab and point to the primary source, so every claim can be verified at origin.

Written and reviewed by

Jack Muncaster · Founder, UK Peptides

Jack founded UK Peptides in Manchester after repeatedly receiving research compounds with missing or recycled paperwork. He is responsible for supplier selection, batch release decisions and the content published in this research library. Every article here is sourced to primary literature and every product page to a signed third-party certificate.

How we write and check this content Report a factual error in this article

More GLP-1 & Incretin Science articles

Popular across the research hub

One flagship guide from every other research category — keep exploring.

← Back to Forum