The stroke gap was 2.1 points; the death gap was 6.2

Two points, six points

The tirzepatide comparison drew 14,118 matched pairs of adults with atrial fibrillation and type 2 diabetes from the TriNetX Global Collaborative Network, an international pool of routine clinical care. At one year the absolute difference in stroke risk was 2.1 percentage points, and the difference in death from any cause was 6.2 percentage points.^[1]

A 2.1-point difference in stroke is a size worth arguing about in a high-risk population. The 6.2 points in death from any cause are both the larger absolute gap and the bolder claim: they say the choice between two glucose-lowering classes moved one-year survival. That is the number that has to pass scrutiny first.^[1]

The plainest account of a survival gap that size is which patients doctors put on tirzepatide. Because treatment was not randomly assigned, the comparison cannot separate the drug from the reasons it was chosen, and matching on measured characteristics cannot balance frailty, functional status or intensity of care that nobody ever wrote down. That is exactly why the authors write that randomized trials are needed before treatment advice changes.^[1]

The same network, a second question

The second study out the same day puts the same infrastructure to work on tuberculosis. Across 143 healthcare organizations and patient data from 2017 to 2025, tuberculosis ran at 0.68 cases per 1,000 person-years among people taking a GLP-1 receptor agonist against 1.67 on sulfonylureas. The hazard ratios came to 0.53 against sulfonylureas, 0.60 against metformin, 0.49 against DPP-4 inhibitors and 0.82 against SGLT2 inhibitors.^[2]

The order of those four numbers carries information by itself. The gap is widest against sulfonylureas and DPP-4 inhibitors and narrowest against SGLT2 inhibitors — the comparator that goes to the patients most like the ones who receive a GLP-1 drug. A ranking that moves with who the comparison group is, more than with the drug under study, carries the fingerprint of residual confounding.^[2]

The absolute scale asks for the same care. Tuberculosis is uncommon in this population, so the difference against sulfonylureas comes to about 1 case per 1,000 person-years. The authors describe the finding as an association and do not claim the drugs prevent infection, and it is not a size that would move a patient's choice of medicine on its own.^[2]

What reaches the patient?

A week ago in this column I wrote about what the measure misses when a surgery candidate's illness is reconstructed from diagnostic codes. The problem here is from the same family: data gathered during routine care brings in more patients than any trial could reach, yet the variables it carries are only part of what decides who receives which drug. The gain is speed; the limit is that whatever a clinician never wrote down cannot be adjusted for either.^{[2], [3]}

For a patient with atrial fibrillation and type 2 diabetes, the honest answer today is this: these two studies say which questions deserve to be asked with random assignment, and not yet which drug prevents a stroke. If the 2.1-point stroke difference in the tirzepatide comparison reappears in a randomized trial in the same population, it becomes a real treatment gain; if it does not, the question of where the 6.2 points in mortality came from is what remains.^[1]