Before any of the numbers, what was measured: not a cognitive test, but whether somebody writing up an appointment noted a problem with memory, attention or orientation. That is a real signal and it is also a record of how often a person was seen and how carefully they were written up — the sort of measurement question that decides what a comparison can be asked to carry.
The setup
Adults with any psychiatric diagnosis in a multicenter US electronic health record, each with at least one clinician-rated cognitive assessment in the year before starting an antidiabetic drug and at least one in the year after. [1] Scores run 0 to 100, where 0 means no signs recorded and 100 means signs across every domain.
Mean age 62.0, 48.4% women. Semaglutide starters were compared against four alternatives: sitagliptin, empagliflozin, glipizide and no antidiabetic treatment at all.
What the four comparisons gave
Against no treatment, semaglutide scored 11.14 to 14.91, a mean ratio of 0.75, 95% CI 0.65 to 0.85, p = 0.00002 after Bonferroni correction. Against glipizide, 0.83. Against empagliflozin, 0.81.
Against sitagliptin: 11.14 to 12.52, a mean ratio of 0.89, 95% CI 0.77 to 1.02, p = 0.50. Not significant.
The weakest arm gives the biggest number
“No antidiabetic treatment” produced the largest ratio, and it is the arm to trust least. Somebody with a psychiatric diagnosis and no antidiabetic prescription differs from somebody just started on one — in how sick they are, how engaged with care, and how often anybody writes anything down about them.
That is the standing problem with a comparison drawn from records rather than built by randomization, and it runs the same way every time: it flatters the treated group, the way three significant differences vanished once the groups were made comparable elsewhere in this section.
Where it held and where it did not
Associations were strongest for memory and for broader cognitive function. Across the disorder groups the effect sizes looked similar, but statistical significance held consistently only in major depression.
The authors’ own closing line asks for randomized trials to establish whether any of this is causal or useful, which is the correct request and worth more than the headline ratio — and it is the same request every observational result on this site ought to end with.
What a reader should take
Not a reason to buy anything. Nobody in this study was taking semaglutide for cognition, and no seller on this site sells it for that.
What is worth keeping is the shape: an outcome assembled from clinical notes, four comparators, and a result that survives three of them and not the closest one. When a finding depends on which comparator you pick, the comparator is the finding — the same lesson as a trial that gave two answers depending on what it was measured against.