Of 34,956 adverse event reports filed in Europe across six GLP-1 receptor agonists between January 2022 and September 2024, semaglutide carried roughly half the per-report rate of thyroid cancer events that tirzepatide did (reporting odds ratio 0.54, 95% confidence interval 0.37 to 0.81, P less than 0.05), Italian researchers wrote in Pharmacological Reports ↗ on May 28. When the analysis was stratified by what indication the drug was prescribed for, the gap disappeared.
EudraVigilance is the European Medicines Agency's pharmacovigilance database. Doctors, patients, and manufacturers can file reports linking a suspected drug to an observed event; the EMA uses the reports to look for safety signals. A reporting odds ratio compares how often event X is reported with drug A versus drug B as a fraction of each drug's total reports. ROR less than 1 means relatively fewer; greater than 1 means relatively more. It is not a measure of true population risk. It is a measure of disproportional reporting.
The Italian team pulled all individual case safety reports listing one of six GLP-1 receptor agonists (semaglutide, liraglutide, exenatide, lixisenatide, dulaglutide, or tirzepatide) as the suspected drug between January 1, 2022 and September 26, 2024. They restricted attention to reports tagged with thyroid cancer-related adverse events. The headline disproportionality finding was the semaglutide ↗ versus tirzepatide ↗ comparison: ROR 0.54, 95% CI 0.37 to 0.81, P less than 0.05. The other within-class comparisons did not reach statistical significance. Most patients filing reports were adult or elderly women. The three most commonly reported system organ classes across all reports were gastrointestinal disorders, general disorders and administration site conditions, and injury, poisoning, and procedural complications.
The sensitivity analysis is the load-bearing footnote. When the team re-ran the analysis stratified by therapeutic indication (diabetes versus obesity), no significant signals remained across stratified therapeutic groups. That means the unstratified ROR 0.54 result is most plausibly driven by the mix of indications in the underlying cohort rather than by a within-class drug difference. Tirzepatide ↗ (marketed as Mounjaro and Zepbound) was more commonly prescribed for diabetes than for obesity in this two-year window; the Zepbound obesity approval landed in late 2023. Semaglutide ↗ carried both diabetes (Ozempic) and obesity (Wegovy) labels for the full window. The patient populations under those labels differ on age, baseline cancer risk, and how aggressively clinicians screen for incidental thyroid findings on imaging. Strata that match on indication match on roughly that demographic axis. The signal vanishes once they are matched.
The broader context the paper sits in is the FDA boxed warning that covers the entire GLP-1 class for medullary thyroid carcinoma and multiple endocrine neoplasia type 2 syndrome. The boxed warning was originally driven by rodent C-cell tumor signals from liraglutide and exenatide preclinical work; the human evidence base for a real thyroid cancer signal in routine use has stayed thin. Large population-based cohort analyses in the last few years have, on balance, not found a clear class-level excess. The current paper does not change that picture. What it does change is the framing of within-class comparisons: a raw disproportionality result that looks like one GLP-1 drug is safer than another should be checked against indication strata before it is read as a safety differential.
For prescribers, the practical takeaway is small. Within-class thyroid safety remains an open question that requires indication-matched cohorts, not raw reporting-odds comparisons, to move. For platform readers, the relevant peptidemodel cards span the class: semaglutide ↗, tirzepatide ↗, liraglutide ↗, exenatide ↗, lixisenatide ↗, and dulaglutide ↗, all on the GLP-1R ↗ target page. Tirzepatide also engages GIPR ↗, which has its own emerging safety literature distinct from GLP-1R.
The cleanest sentence in the paper is also the most modest one: the authors flag that their findings should be interpreted with caution given the inherent limitations of pharmacovigilance databases. That is the only line the casual reader needs to absorb. The disproportionality moved with the strata, and the strata are the story.