One comparison in this paper is printed three times, in two directions. That is the part worth your attention, and it sits inside a design that was otherwise careful.
What was compared with what
Adults with type 2 diabetes who started a GLP-1 drug were matched one-to-one against adults who started insulin, metformin, a DPP-4 inhibitor, an SGLT2 inhibitor, a sulfonylurea or a thiazolidinedione. [1] Matching used 39 baseline variables covering demographics, conditions, procedures and other prescriptions.
The outcome was a pancreatic cancer diagnosis recorded between 365 and 7,300 days after the first prescription. That first year is deliberately discarded, because pancreatic cancer often causes diabetes months before anyone finds the tumor. Without the gap, the drug someone was prescribed for their new diabetes would look like the cause of the cancer that produced it.
Four lower, one level
Against metformin the hazard ratio was 1.39, interval 1.16 to 1.66, in the comparator’s direction — so the metformin group carried the higher hazard. Sulfonylureas gave 1.37, DPP-4 inhibitors 1.31, thiazolidinediones 1.30 with an interval of 1.001 to 1.678. SGLT2 inhibitors gave 1.08, and that interval crosses one.
Read one at a time, each is a separate observational study with its own matched population. We have written about what that does to an answer and will not repeat it here. What is new is the sixth row.
How large are these differences
Small. The absolute risk differences run from 0.04 to 0.15 percentage points, on underlying risks of roughly one in a thousand.
A hazard ratio of 1.39 sounds like a third more cancer. In this data it is about four extra diagnoses per ten thousand people over a period that can run twenty years. Ratios and counts are two different sentences, and a large ratio on a small base is the easiest number on a page to misread.
What matching does not fix
Nobody here was randomized. People put on insulin have usually had diabetes longer and controlled it worse, and 39 covariates do not erase that. The comparison against insulin is the one most likely to be carrying it, which makes the contradiction in that row doubly awkward.
The record network also only knows what was coded. A diagnosis appears when somebody wrote it down, and so do the conditions used to match on — the same limitation behind the other large matched-record study on cancer.
What to take from it
No signal that GLP-1 drugs cause pancreatic cancer. That is the useful finding, and it is a negative one.
Anything stronger runs into the comparator problem. “Lower risk” is not a property a drug has on its own, and a difference of a fraction of a percentage point is not something a person can feel either way.