Peptide basics

How to read a sex split in a trial result without over-reading it

This page contains no dose, schedule or route and makes no claim about any compound. It is a method page: how to judge whether a sex-split result you have been shown supports the conclusion attached to it.

The mistake the checklist prevents

A trial reports a result in women and a result in men. One looks bigger. The natural next sentence — that the drug works better in women — is a claim about an interaction between sex and treatment, and it does not follow from two separate within-group estimates.¹

The reason is that each estimate carries its own uncertainty, and the smaller group carries more of it. Two intervals that overlap substantially can still produce point estimates that look far apart, especially when one group is a third the size of the other.¹⁴

The five questions

The NEJM methods paper on subgroup reporting sets out what an author should provide and what a reader should look for. Reduced to questions you can ask of any sex-split figure, it comes to five.¹

1. Was the subgroup analysis pre-specified in the protocol or statistical analysis plan, before the data were unblinded?
2. Was a formal test of interaction reported, rather than only two separate estimates?
3. How many subgroups were examined in total? Many comparisons raise the chance of a spurious one.
4. Was the subgroup variable measured at baseline, and was it used to stratify randomisation?
5. Has the same direction of effect been seen in other trials, or is this the only place it appears?¹²

The BMJ review that formalised these criteria applied them to 207 randomised trials published in core clinical journals that reported subgroup analyses. Sixty-four made a subgroup claim for the primary outcome.²

Of those 64 claims, 26 (41%) had clearly pre-specified their hypotheses, 4 (6%) had correctly pre-specified the direction, 13 (20%) had used the subgroup variable as a stratification factor at randomisation, and 6 (9%) reported a test of interaction that proved statistically significant.²

Running the checklist on the GLP-1 sex splits

The most widely repeated sex split in this literature is the STEP 1 comparison reported in a 2023 review: an estimated treatment difference in weight reduction of −14.0% in women against −8.0% in men. Two within-group estimates, presented side by side.⁵

Question one: we are not aware of documentation that sex was a pre-specified subgroup with a stated directional hypothesis in that trial. Question two: no formal interaction test accompanies the figures as usually quoted. Question four: STEP 1's registry record does not indicate sex was used as a randomisation stratification factor.⁴⁵

Question five is where the claim does better. A 2025 meta-analysis pooled 14 randomised trials reporting weight change by sex across five compounds and found a consistent direction: females lost more weight than males, with a pooled mean difference of 1.04 kg and 1.69 percentage points.³

Replication across trials and compounds is the strongest thing this claim has going for it. It is also the criterion the BMJ review found satisfied least often in practice, so it is worth weighting.²³

What a pooled estimate adds, and what it does not

Pooling across trials increases the number of participants behind an estimate and shows whether the direction is stable. Both are genuine gains over a single trial's split.³

It does not repair the underlying design. If none of the pooled trials pre-specified sex or tested an interaction, the pooled result is a summary of many weak analyses rather than one strong one. The 2025 authors flagged this directly, noting that incorporating post-hoc analyses raised the risk of bias and that many included studies lacked granular sex-stratified data.³

It also cannot tell you anything about an individual. A one-kilogram average difference between two large groups is compatible with enormous overlap between them.³

Semaglutide's sex-split figures are the ones most often reproduced. Before repeating one, check whether the source reports an interaction test or only two separate estimates — in the widely circulated version, it is the latter.

Common questions