The largest cohort points the other way, and the reason to be careful with it is not the direction but the missing sizes. Among 63,215 people who already carried a neuropsychiatric diagnosis, those who reached a higher semaglutide dose in their first two years had fewer events over the following two [1]. Seven outcome categories moved, and the abstract reports every one of them as a P-value with no hazard ratio, no interval and no rate in either group. On a site whose argument is that a number needs a size, that absence is the finding, and a dose here is also a price that usually climbs with it.
What was measured
The cohort was observational and tracked 24 incident outcomes after treatment started. Propensity-matched against metformin, SGLT2 inhibitors and DPP-4 inhibitors, semaglutide came out lower on neuropsychiatric event risk over two years. The second analysis is the one the question turns on. Within the semaglutide patients, the authors took the highest dose a person attained across their first two years, then counted events over the following two.
That separation of windows is a landmark design, and it does real work. It stops a later event from explaining an earlier dose, which is the commonest way a dose-response finding in records data turns out to be running backwards.
Seven findings and no sizes
Substance-related disorders, mood disorders, anxiety and stress disorders and central nervous system atrophies all came in at P < 0.001. Neuromuscular disorders sat at P = 0.013, eating, sleep and behavioral disorders at P = 0.022, and personality and impulse-control disorders at P = 0.028. Dementia and degenerative disease showed no dose difference at all, P = 0.15, which the authors note agrees with earlier trials.
A P-value says an association is unlikely to be chance. It says nothing about whether the difference is one event per thousand or one per ten, and that gap is the entire distance between a finding and a decision anybody could act on.
What the class-level evidence says
A Swedish national cohort followed 95,490 people with depression or anxiety who were taking an antidiabetic drug, of whom 22,480 used a GLP-1, and analyzed each molecule separately [2]. Semaglutide carried an adjusted hazard ratio of 0.58 for worsening mental illness, 95% CI 0.51 to 0.65, and liraglutide 0.82, 95% CI 0.76 to 0.89. Exenatide and dulaglutide both came in at 1.01, with intervals crossing no effect.
That study measures molecules rather than doses, and its outcomes are registry events like psychiatric hospitalization rather than questionnaire scores. Two drugs in one class showing nothing is a warning against reading any of this as a class property, which is the point made at length in four drugs in one class. Tirzepatide was in neither analysis.
What a reader should do with it
Not titrate. Nothing here is a treatment claim for anyone with a psychiatric diagnosis, nobody was randomized to a dose, and the decision belongs to a prescriber who knows the rest of the medication list. The reporting data on the specific fear is separate and sits in reports against risk.
What is useful is the direction. Public concern about mood on these drugs has run the opposite way, and two large cohorts pointing at lower risk is worth knowing even at observational strength. The price question is unchanged: what a maintenance dose costs depends on the dose people actually stay on and on how many are still paying a year later.