Two Studies, One Comparison, Two Answers

Two real-world studies published this year compared semaglutide against tirzepatide on cardiovascular outcomes. One found no meaningful difference. The other found semaglutide reduced major cardiovascular events by 29%.^[1][2] Both used large databases and careful matching. Working out why they disagree is more useful than either answer.

The two results

Reason one: they studied different people

This is the most straightforward explanation and possibly sufficient on its own. One study looked at people with type 2 diabetes. The other looked at people with overweight or obesity and established atherosclerotic cardiovascular disease but no diabetes.

Those groups have different baseline risks, different reasons for being prescribed either drug, and different competing causes of cardiovascular events. There is no rule that two drugs must rank identically in both.

Reason two: one result is marginal

The p-values in the claims study are 0.046 and 0.040. Those clear the conventional threshold by a hair. A finding at p=0.046 is one that would frequently not replicate, and it should be read as “possibly something” rather than as a demonstration.

Reason three, and the one to learn from

The claims study also ran a per-protocol analysis, censoring patients when they had a gap in therapy of more than 30 days. In that analysis the hazard ratio moved from 0.71 to 0.43, and from 0.78 to 0.57 — a far more dramatic effect, presented as the stronger finding.

An effect that grows sharply once you exclude the people who stopped is a warning sign, not a confirmation.

Excluding people who discontinue removes a specific kind of patient: the ones who could not tolerate the drug, could not afford it, got sicker, or were hospitalized. Those are not random departures — they are systematically people with worse trajectories. Removing them makes any drug look better, and it makes the drug with better persistence look better still.

We know persistence differs sharply between these drugs and formulations, because a UK cohort measured it: discontinuation ran from 36.9% to 55.8% depending on the drug. A comparison that censors on discontinuation is partly measuring that.

Which to believe

The trial-emulation study has the stronger methodology, and it is worth explaining why. Before extending to a broader population, it first reproduced the reference randomized trials — checking that its methods recovered the known answers on every individual endpoint except all-cause mortality in one trial, and only then applying those methods to new questions.^[1]

That is calibration, and it is rare. Most observational studies ask you to trust their design; this one demonstrated it against a known answer first. Its conclusion — comparable cardiovascular benefit between the two drugs in clinical practice — is the one we would weight more heavily.