An incidence rate ratio of 0.87 for major adverse cardiovascular events, drawn from 99,599 patients across 21 randomized trials, graded high-certainty, number-needed-to-treat of 66. That is not a signal. That is an endpoint.
Now hold it against a headline you also saw this month: GLP-1s "linked to lower breast cancer," odds ratio 0.695, from a database of 111,646 women.
Both claims traveled the internet at the same velocity. Both sound like the same kind of news. They are not. They sit on opposite sides of an evidentiary chasm that the coverage keeps papering over, and the honest work this week is naming which side each one is on.
Both claims traveled the internet at the same velocity. They sit on opposite sides of an evidentiary chasm that the coverage keeps papering over.
Last week this column stayed inside single organs — GLP-1 agonists in heart failure and Parkinson's. This week we widen to the harder problem: what happens when one drug class plausibly touches the heart, the kidney, and the tumor at once, and the evidence for each claim was built by completely different machinery.
Start with the machinery that works. The JACC meta-analysis pooled eight different agents — lixisenatide through tirzepatide — over a mean 2.4 years of follow-up, and reported conclusive, high-certainty reductions in all-cause death (IRR 0.88, NNT 121), cardiovascular death (IRR 0.87, NNT 170), and MACE (IRR 0.87, NNT 66).
These are randomized comparisons against placebo or control, run through GRADE and trial sequential analysis, the two tools whose entire job is to tell you when a conclusion is durable rather than fragile.
The mechanistic story is consistent with an effect that outruns weight: benefits appear early after initiation, before much fat is lost, plausibly through endothelial function, plaque stability, blood pressure, and epicardial fat. That is why the field says "beyond glycemic control" and means it.
Bank one more number from that same analysis, because it becomes load-bearing later: no difference in neoplasm between groups. In the cleanest dataset we have, GLP-1 RAs did not raise or lower cancer. They were neutral.
Now walk into the seductive middle, where real-world evidence lives and the numbers get bigger and softer at the same time.
The new tirzepatide study is the exemplar. Analyzing 52,971 US patients over 40 with type 2 diabetes and established atherosclerotic disease, it compared tirzepatide against sitagliptin — chosen precisely because sitagliptin has no documented cardiovascular effect, making it a clean placebo proxy.
The result: MACE of 2.9% versus 4.4% at one year, a 32% relative reduction. Larger than anything in the randomized pool.
But read the fine print the authors themselves wrote. The benefit "depends substantially on outcome definition" — myocardial infarction was lower, ischemic stroke showed no meaningful difference. And 21% of patients were taking SGLT-2 inhibitors that the analysis never accounted for.
The editorialists were blunt: consistency is "encouraging," but "it does not prove incremental benefit over optimized SGLT-2 therapy," because background treatment was not randomized, adherence was incompletely observed, and no treatment-by-SGLT-2 interaction was shown. Their closing line is the one to keep: "The signal is compelling; the causal and clinical placement questions remain open."
Here is the pivot the coverage misses. This is a good study. Real-world evidence, "trial-anchored" and propensity-weighted, is a legitimate way to inform a decision when you already have a wall of randomized data pointing the same direction. It fills in the placement question for cardiology.
The exact same methodology, pointed at cancer, produces numbers you cannot lean on — because there is no randomized wall behind them.
The breast cancer EHR study makes this vivid. It identified 111,646 women aged 45 to 80 with a BMI of at least 25 and a documented breast imaging outcome, then propensity-matched on age, race, ethnicity, BMI, breast density, and diabetes history — a study that, per the PubMed report, measured breast cancer detection rather than survival or recurrence. GLP-1 exposure was associated with lower breast cancer incidence — OR 0.649 unmatched, 0.695 matched, p below 0.0001.
Impressive until you ask what it is measuring. Detection at imaging, coded from records, not adjudicated tumor biology. First prescription before the exam date. And the whole edifice rests on propensity scores, which can only balance the confounders you thought to measure.
A separate oncology dataset — reportedly 841,831 women in a JAMA Network Open study, with GLP-1s associated with better survival and recurrence outcomes among women who already have breast cancer with obesity or type 2 diabetes — carries the same architecture and the same tell. When the comparison was run against SGLT2 inhibitors, another modern metabolic drug class, the signal weakened.
That weakening is the whole story. It suggests the apparent benefit may not be GLP-1-specific at all, but a generic dividend of better metabolic control — the kind you might get from any drug that improves the same underlying physiology.
Put the three datasets in one line and the split resolves itself. Randomized trials: cancer-neutral. Retrospective EHR: cancer-protective, but the protection dissolves against a metabolic comparator. The most parsimonious reading is that the oncology "signal" is confounding wearing a lab coat.
None of this means the cancer question is dead. The biology is real — obesity reportedly worsens prognosis (recurrence and mortality) by 35 to 40%, per the MDPI review, and GLP-1s move the hormonal, inflammatory, and insulin pathways that connect the two. That is exactly why it deserves prospective randomized testing, which every serious author here explicitly calls for.
Hypothesis-generating is not an insult. It is a job title. The error is promoting the hypothesis to a recommendation because the odds ratio was large and the p-value small.
Which brings us to the tell that makes this whole moment legible: the field is convening measurement frameworks before the causal questions are answered, not after.
ISPOR is running a webinar this month titled "Whole Health Evaluation of Multi-System Risk: GLP-1 Insights," pitching a framework that evaluates a therapy across cardiometabolic, renal, and behavioral domains at once — explicitly because traditional evaluation "focus[es] on isolated endpoints measured under controlled conditions" and misses how a drug acts as an "upstream modifier of multi-system risk."
Read that as a confession dressed as an agenda. When a drug class plausibly touches heart, kidney, and cancer simultaneously, and its real-world effect is "shaped by treatment dynamics and patient context" — adherence, discontinuation, access — you need new scaffolding to tell weight-mediated effects from weight-independent ones, and to separate drug from confounder.
The frameworks are being built precisely because the current toolkit cannot yet make that separation. They are honest about their own immaturity. The point is to distinguish efficacy from effectiveness, not to declare victory across every organ.
So here is the discipline the coverage owes its readers.
Let the cardiovascular story graduate. Two tiers of evidence — high-certainty randomized meta-analysis plus corroborating real-world data — now point the same way. MACE reduction is practice-relevant. An NNT of 66 is a number a clinician can act on.
Quarantine the oncology signal as hypothesis-generating. Not because the biology is implausible, but because every eye-catching cancer number so far is a retrospective association, and the one place with randomization found no difference in neoplasm at all.
The two headlines you saw this month were never equivalent. One reports an endpoint. The other reports a reason to run a trial. Treat them the same and you will make a worse decision than the evidence supports — in either direction.
The most useful thing a reader can carry out of this issue is the habit of asking, before the excitement lands, one question: randomized, or retrospective? The answer sorts the news every time.