167 - Gary Taubes: Bad science and challenging the conventional wisdom of obesity

Taubes describes trying Atkins himself during reporting the New York Times Magazine piece: lost 25 pounds in six weeks. He notes that the first five clinical trials comparing Atkins to low-fat calorie-restricted diets — including Eric Westman's first trial at Duke and a VA hospital group — all showed the Atkins arm losing more weight with better cardiovascular risk factor profiles. This was a double refutation: it contradicted the caloric-excess model of obesity (subjects could eat ad libitum and still lose more weight) and it contradicted the dietary-fat/heart-disease hypothesis (higher fat intake did not worsen risk factors).

Taubes references Virta Health as an example of capitalism filling the gap left by the medical establishment's slow adoption of carbohydrate restriction: 'you have operations like virta health which is doing very well advocating using nutritional ketosis to treat type 2 diabetes.' Virta's model is the commercial translation of the research Taubes had been advocating for through NuSI — and its survival and growth constitutes real-world evidence that the intervention works at scale.

Taubes describes the correct pre-publication protocol from physics: you do not announce a discovery until you have gone to 20 seminars and asked every physicist you can find to explain how you screwed up. Only when nobody can tell you do you publish, still with a question mark in the title. The contrast with nutrition: the Nurses Health Study published associations between diet and heart disease / cancer as if they were causal, based on questionnaire data from a non-random sample, without measuring household income, without randomization, with no attempt to enumerate the 'Altarelli cocktail' of alternative explanations.

Taubes's most vivid illustration: Georgia Tech chemists held neutron detectors over cold fusion cells, detectors went off, they held a press conference. A physics colleague pointed out at the press conference that the detectors were humidity-sensitive and the cells were bubbling. They ran the water-only control within 12 hours of the press conference and immediately retracted. The failure was not fraud — they genuinely had not thought to run the control. The parallel in nutrition: every food-frequency questionnaire study faces the healthy-user bias, but the Nurses Health Study and most large observational studies did not adequately measure it.

Taubes credits this protocol to the physics community's culture — and contrasts it with what Rubia at CERN did (went public at a Washington conference before doing the background analysis) and what Pons and Fleischmann did (issued a press release because they feared a competing group would steal their discovery). The competitive pressure to publish first is the single most common trigger for the pathological science flip from rigor to advocacy.

Taubes notes that insulin measurement was impossible before the Yalow-Berson radioimmunoassay of 1960 — which is precisely why the 1940s and 1950s fat-metabolism literature could not directly test the insulin hypothesis. Now that measurement is routine and cheap, fasting insulin arguably should be standard of care alongside fasting glucose. The carbohydrate-insulin model predicts that patients who are obese with high fasting insulin will respond better to carbohydrate restriction than to calorie restriction — testable in individual clinical practice.

Taubes draws an analogy to COVID vaccine research mobilization: 'think about the kind of effort that went into COVID — the amount of money that was spent on research to prevent this disease from killing what one tenth of the number of people who die from chronic diseases related to obesity and diabetes every year.' The obesity and diabetes epidemics are not stemmed; the conventional wisdom's failure is the only signal strong enough to force a paradigm examination. 'Step one is saying we failed.'

The mechanism Taubes proposes: a VMH lesion triggers hyperinsulinemia; elevated insulin signals fat tissue to take up and store fatty acids, simultaneously inhibiting lipolysis and fatty-acid oxidation in muscle (via the malonyl-CoA pathway), creating a peripheral starvation state that explains both the gasping hunger and the fat gain in the same model. Brobeck never considered this because he entered the field having already accepted Newberg's 'eating too much' framework and interpreted the pair-feeding anomalies (three of twelve animals got fat despite being pair-fed) as experimental error rather than evidence against his hypothesis.

Taubes is explicit that when he wrote the NYT Magazine article (circa 2001) he still had a line saying 'overweight is caused by taking in more calories' — which he now considers 'wrong and meaningless.' His position evolved through writing Good Calories Bad Calories: the first third deconstructs the dietary-fat hypothesis, the second develops the carbohydrate-centric hypothesis from British nutritionists and insulin-resistance researchers, and the third examines obesity as a hormonal fat-storage disorder. The diabetes and heart disease data were already pointing the same direction by the time those five clinical trials were published.

The canonical Langmuir failure mode, per Taubes: researcher sees signal → goes public before doing background analysis → switches from 'trying to prove myself wrong' to 'collecting evidence that I'm right' → surrounding themselves with people who confirm the belief, dismissing skeptics. Rubia's group did this at CERN; Pons and Fleischmann did this with cold fusion; the Georgia Tech chemists who announced cold fusion neutron detection at a press conference and retracted within 12 hours were humidity-sensitive detectors over bubbling cells — the classic failure to identify your control. The nutrition parallel: the Nurses Health Study never measured household income, a first-order confound in virtually all its diet-disease associations.

The power-line cancer example: people living near power lines may have lower socioeconomic status (power lines are unsightly, noise-making; those who can afford to move, do). That's a first-order confound for cancer rates. The Nurses Health Study — the most famous cohort study in America — never measured household income. When physicists say 'five sigma,' they're saying 'we've thought of every first, second, and third-order alternative explanation and left a huge buffer for the ones we haven't thought of yet.' Epidemiologists claiming causality from a relative risk of 1.3 at p=0.05 are making a claim physicists would not publish until they'd ruled out vastly more of the background.

The dbdb mouse is the same genetic mutation on a different background strain: both are hyperinsulinemic from weaning onward, but the dbdb background strain cannot sustain the hyperinsulinemia — its pancreas fails and produces frank diabetes. The obob background strain keeps pumping insulin and stays obese without becoming overtly diabetic. Taubes: this is direct evidence that hyperinsulinemia (not hyperphagia) is the primary lesion, and the strain difference just determines whether the animal develops obesity alone or obesity plus type 2 diabetes.

Kevin Hall, the NIH researcher brought in as a 'physicist's perspective outsider' to run the energy-balance arm, initially appeared to share Taubes's skepticism of conventional wisdom. Taubes argues that once Hall started interpreting the energy-balance consortium data as supportive of his own computational model and as refuting the carbohydrate-insulin model, confirmation bias locked in — the same trajectory he had documented in Rubia and in Pons and Fleischmann.