Healthy Skepticism
Each paper takes one claim and works it through by counting, with the sources named. They were written alongside the course and stand on their own. Free to read, print and pass along. New papers are announced first in the newsletter.
Relative reporting
Four numbers come out of one 2×2 table, and which one you report decides how impressive the result sounds. Worked through twice, for atorvastatin (Lipitor) and for the COVID-19 vaccines, with the traps to flag in any risk claim and a teaching example at the back.
The Bernoulli fallacy
CERN announced the Higgs boson at a fluke probability of about one in 3.5 million. A medical journal will publish at one in twenty. Both results get called significant. The paper walks the gap between the two questions, P(data | the null is true) and P(the hypothesis | data), and what it takes to get from one to the other.
The syndrome trap
A syndrome is a checklist with a name, not a disease, but the label gets talked about as if it had one mechanism and one treatment. Metabolic syndrome names sixteen different clinical pictures needing different medicine. Treating its checklist numbers failed outright, and the one drug class that helps the whole cluster arrived by accident, out of diabetes trials, because it hit an upstream mechanism the checklist never named. The same trap stretches across the autism spectrum, where one diagnosis spans a non-verbal child and a chief executive, and across ADHD, where two people can share the diagnosis and not one symptom; the new "multi-dimensional spectrum" is the old concept of a syndrome under a fresh name. Ends with the three questions that undo the trap.
Case study · Faults 1, 2, 4 and 5
Nicotine pouches are sold as a focus aid, a brain protector, and a safer habit. One wellness claim turns out to be four claims, and each rests on a different fault: who uses it read as what it does, a Parkinson's correlation with the arrow running backward, a small memory trial whose two-year follow-up found nothing, and "safer than smoking" with the comparison left unsaid. A case study in applying the faults to a real claim, written for readers who have not taken the course.
Half the people on the reduced dose of Eliquis (apixaban) do not meet the label's criteria for it, and no trial has ever compared the two doses directly. The paper measures the drug against the alternative patients actually face, not taking it, works out where the 2.5 mg rule came from, and builds a Bayesian model that turns the published trials into personal numbers, with honest ranges where the evidence runs out.
The paper travels with a kit for patients and their cardiologists: a short patient guide, a technical paper and one-page model card for the doctor, and a locked spreadsheet the two of them run together in a visit. The kit, explained piece by piece.
Body weight is regulated by hormones, not by resolve, and GLP-1 signalling runs weak in people carrying extra weight, so the drug supplies an input the body is not producing at strength. Works through what the hormone measurements show a year after a diet, why the concept of cheating does not apply, and what it means that the weight returns if you stop.
Risk of vascular death climbs smoothly with usual pressure, with no safe ledge down to about 115 over 75, which is exactly what a cutoff at 130 or 140 hides. The paper separates what systolic and diastolic each tell you, then builds a small Bayesian network that takes age, sex and the two readings and returns a ten-year probability.
How the BMI curves used in the course are built: shape from a pooled analysis of 239 studies, level from a life table, context from the population distribution. The WHO bands only shade the figures. They never enter the calculation, because the bands are a lossy summary of a smooth signal rather than biological boundaries.