GLP-1s and Worsening Mental Illness

A Swedish cohort followed people with depression and anxiety who took GLP-1 drugs, and found something more interesting than a simple answer. Semaglutide was associated with a 42% lower risk of worsening mental illness. Liraglutide, 18% lower. Exenatide and dulaglutide, nothing at all.^[1] Four drugs, one class, two very different results.

What was found

For semaglutide the effect appeared across all three specific outcomes: worsening depression at 0.56, worsening anxiety at 0.62, and worsening substance use disorder at 0.53. Liraglutide moved only depression, at 0.74. And across the class as a whole, self-harm was lower — 0.56, with a confidence interval from 0.34 to 0.92.^[1]

Why the split between drugs is the finding

We have now seen this shape twice. In a meta-analysis of atrial fibrillation, the class was neutral and the signal appeared specific to semaglutide, which we covered in reading a 42% headline. Here again, two drugs move and two do not.

An effect that lands on some members of a drug class and not others is either real pharmacology or a description of who receives which drug.

Both explanations are live and they are not equally likely in an observational study. Semaglutide is the newest and most effective of the four, which means it is prescribed to different people than exenatide — more often younger, better insured, more engaged with their care, more likely to be losing substantial weight. Every one of those independently predicts a better mental-health trajectory.

There is also a straightforward mediating route that has nothing to do with the brain: semaglutide produces far more weight loss than exenatide, and losing a great deal of weight changes mobility, sleep, pain and how people are treated by others. A drug that improves depression through those channels is still helping; it is just not evidence of a direct psychiatric mechanism.

The self-harm result, and where it sits

The class-wide self-harm finding matters because it speaks to the question this drug class has been most publicly accused over. A hazard ratio of 0.56 points in the opposite direction from harm, in a population selected for existing depression and anxiety — which is exactly the group any risk would show up in first.

Set that against the spontaneous-reporting signal we covered in the FAERS analysis, which pointed the other way. These are not equally strong designs: a matched national cohort with adjustment beats voluntary reports with no denominator, and the regulatory picture is covered in what happened with the suicidality signal.