Adding a Warning Needs Suspicion; Removing One Needs Evidence

Signals are opened easily and closed slowly, on purpose

The threshold for taking a safety question seriously is deliberately low. Spontaneous reports arrive unverified and uncontrolled, and a regulator that waited for proof before investigating would be too slow to be useful. The threshold for declaring a question settled is deliberately high, because the cost of wrongly removing a warning is borne by patients while the cost of wrongly keeping one is mostly borne by the drug. The result is an asymmetry that looks like inconsistency and is not: the same body can say a warning is justified on suspicion and, later, that removing it requires far more than suspicion to the contrary.

The sequence, with dates

The FDA opened a formal investigation in July 2023 after postmarketing reports of suicidal ideation and behaviour in people taking GLP-1 receptor agonists. Its preliminary Drug Safety Communication came in January 2024 and reported no clear evidence of causation - while noting that only a small number of cases had been observed in individual trials, so the risk estimate was uncertain. In April 2024 the European Medicines Agency's Pharmacovigilance Risk Assessment Committee concluded that the available evidence did not support a causal association for dulaglutide, exenatide, liraglutide, lixisenatide and semaglutide, having reviewed non-clinical studies, clinical trials, post-marketing surveillance and additional studies including one the agency conducted itself. On 13 January 2026 the FDA requested that manufacturers remove the suicidal behaviour and ideation warning from labelling for Saxenda, Wegovy and Zepbound.

The two regulators did different things, and the difference matters

This is where most summaries collapse the story. The PRAC's April 2024 conclusion was that no update to the product information was warranted - meaning it declined to add anything, and left existing labelling as it stood. The FDA's January 2026 action was to request removal of a warning that was already there. Declining to add and choosing to subtract are different regulatory acts requiring different amounts of evidence, and the twenty-one months between them is roughly what it took to assemble the second.

What made the removal possible

Not more reports, and not more observational studies. The FDA's meta-analysis pooled 91 placebo-controlled clinical trials from across GLP-1 receptor agonist development programmes: 107,910 participants, 60,338 on a GLP-1 receptor agonist and 47,572 on placebo, examining suicidal behaviour and ideation along with other psychiatric adverse events including anxiety, depression, irritability and psychosis. Randomised comparison addresses the central confounder directly - that people prescribed these drugs differ systematically from those who are not, including in baseline psychiatric risk - and a hundred thousand participants supplies the precision that individual trials, each containing only a handful of events, could not.

What the closure does and does not say

It says that across a very large randomised dataset, no increased risk of these outcomes was detected. It does not say the risk is zero, because no study design establishes that; a large randomised analysis places an upper bound on how big an effect could have gone unnoticed, which is the strongest available form of the claim. It also does not extend to populations the trials excluded, to durations longer than the trials ran, or to compounds not in the analysis. And it says nothing about whether any individual person should stop or continue a prescription, which is a clinical question.

Why this case is worth keeping as a template

Most safety signals do not resolve. They persist as unquantified concerns, attached to labels, cited in coverage, and never formally closed, because the study that would close them is expensive and nobody's incentive is to run it. This one closed, in public, with dated regulatory documents on both sides of the Atlantic and a stated evidence base. When the next signal in this class arises - and it will - the useful questions are the ones this case answers: who opened it and when, what evidence would close it, has anyone assembled that evidence, and has any regulator acted on it.

Quick reference