GLP-1 drugs were associated with a 41 percent higher risk of thyroid cancer in a propensity-matched cohort of 144,410 users versus the same number of non-users. Waited three years after the prescription, and the signal was gone. Every other cancer site that moved, moved in the protective direction, and those reductions held up against active comparators.
The numbers come from Endocrine Practice on May 15 ↗, a TriNetX retrospective cohort of adults with type 2 diabetes mellitus between 2015 and 2023, 1:1 propensity-score matched. All-cause mortality was lower in the GLP-1 RA arm (hazard ratio 0.693, meaning death was about 31 percent less likely per person-year of follow-up). Overall cancer incidence was 9.5 percent lower (HR 0.905). The non-endocrine pattern was clean: digestive cancers down 29 percent (HR 0.710), respiratory cancers down 31 percent (HR 0.691), with consistent reductions reported for oral, female genital, and central nervous system sites.
Then the thyroid line. Initial HR 1.411, meaning GLP-1 RA users were diagnosed with thyroid gland cancer about 41 percent more often than matched non-users over the cohort window. That is a real number, and it lands in the middle of a fifteen-year debate. The first GLP-1 RA boxed warning, on liraglutide in 2010, was driven by C-cell hyperplasia and medullary thyroid carcinoma in rodents at high doses, the receptor for which is expressed on rodent C cells at much higher density than in humans. Observational cohorts have been pulling in both directions ever since, and the warning persists on semaglutide ↗ and liraglutide ↗ labels.
What landmark analysis does
Landmark analysis cuts the early follow-up window. The reasoning: when a patient starts a new prescription, the surrounding workup (extra labs, ultrasound, a thyroid panel ordered alongside the HbA1c chase, an incidental nodule on an unrelated CT) finds cancers that were already there. Counting those as drug-attributable inflates the early hazard ratio. Wait three years, restart the clock, and you measure the rate among people who survived the workup window without a diagnosis. If the drug actually causes new tumors, the hazard ratio should hold or grow. If the early hazard was a detection artifact, it collapses.
The thyroid hazard collapsed. The authors report the 3-year landmark estimate was no longer statistically significant. Their reading: surveillance bias rather than a true carcinogenic effect.
Why the non-endocrine numbers are stronger
The methodological detail that matters here is the active-comparator analysis. Comparing GLP-1 RA users to non-users alone is a weak design because the non-users are a heterogeneous mix of better- and worse-controlled diabetics, healthy avoiders, and people on other agents. The paper also ran the comparison against DPP-4 inhibitor users and against SGLT2 inhibitor users, both of which select for patients similar enough to the GLP-1 RA group on diabetes severity and care intensity to neutralize most of that confounding. The protective findings were "broadly consistent" across those comparators. The thyroid signal in the active-comparator framing got the same treatment from the landmark analysis.
A reasonable summary: the analysis fails to find evidence of carcinogenicity at sites that prescribers and the FDA have been worrying about, and finds protective signals at digestive and respiratory sites that survive against active comparators rather than only against placebo or against no diabetes drug at all. The breast cancer pattern the section reported on May 11 ↗, where the GLP-1 RA protection in JAMA Network Open vanished when matched against SGLT2 inhibitors, is the counter-example. There, the metabolic class shared the benefit. Here, the comparators include SGLT2 inhibitors and the GLP-1 RA signal still holds for digestive and respiratory cancers. Two different cancer-site mechanisms, two different inference patterns, with the SGLT2 active-comparator switch being the diagnostic.
What this leaves open
Three things still owe answers. The rare cancers (medullary thyroid carcinoma specifically, not the papillary thyroid cancers that dominate incidence) need their own follow-up because the divergent-malignancy paper aggregates "thyroid gland cancer" without histology breakdown. The 5-to-10-year horizon is still under-sampled in real-world data. And the question of whether GLP-1 RA exposure during cancer surveillance changes which lesions get caught earlier (a clinical good) or biases the detection statistics in a way that obscures a smaller true risk is now formally on the table for follow-up studies designed to test it directly.
For now, the practical change is that the thyroid line in the prescriber consent conversation has weaker observational support than it did before the landmark analysis ran. The boxed warning will not move on this paper. The literature it now has to compete with grew.