Ten safety outcomes were tested across 13 hospitals, and several came back elevated [1]. Before any of the numbers, the comparator deserves attention, because it decides what the numbers can mean. Everyone in the control group was a matched person who did not start a GLP-1 — and the reasons a person starts a weight drug are not the reasons another person does not — the comparator problem this site set out in how the comparator decides the depression answer.
For semaglutide, psychiatric disorders overall came in at a hazard ratio of 2.02 (95% CI 1.30 to 3.14), anxiety at 2.39 (1.22 to 4.71), depressive disorder at 3.42 (1.51 to 7.74), gastrointestinal dysmotility or obstruction at 3.91 (1.42 to 10.82) and vision impairment at 1.58 (1.04 to 2.41). For liraglutide, with a larger sample, psychiatric disorders were 1.66 (1.36 to 2.03), hepatic impairment 1.46 (1.16 to 1.83), pancreatitis, cholangitis or cholelithiasis 1.40 (1.18 to 1.68) and vision impairment 1.34 (1.16 to 1.55).
The comparator problem is worth stating precisely rather than waving at. Someone who starts a weight-management drug has usually just seen a doctor, is being monitored, and is under a degree of distress about their weight. Their matched non-initiator is not necessarily doing any of those things. Higher recorded rates of anxiety and depression in a group with more clinical contact can reflect more diagnosis, not more disease. Nothing in a matched design removes that.
This is exactly the contrast with the cohort this site covered in the study that tested its own bias, which used an active comparator — people taking a different drug — precisely so that both groups were under care. That choice does not make this study wrong. It makes the two studies answer different questions.
The authors add their own caveat: semaglutide-associated vision impairment and liraglutide-associated hepatic impairment lost significance in some sensitivity analyses. The vision signal in particular joins a literature this site has followed in the sudden vision loss reports, where the same pattern recurs: a real signal, repeatedly detected, never yet cleanly sized.