A Discontinuity Is Not an Experiment

On 26 September 1953 Britain de-rationed sugar, and adult intake roughly doubled — about 40 g/day to 80 g/day — more or less overnight. That single administrative date has become one of the most productive instruments in modern epidemiology, and it has now been worked hard enough that it has stopped measuring sugar. A Europe PMC sweep turns up at least thirteen published papers plus a run of preprints, nearly all drawing on the same ~64,000 UK Biobank participants born 1951–1956, all keyed to that one cutoff. The original — Gracner, Boone and Gertler in Science (2024) — found early-life rationing cut type 2 diabetes by ~35% and hypertension by ~20%: modest magnitudes, a metabolically coherent story. Since then the same discontinuity has been reported to protect against cardiovascular disease, heart failure, fatty liver disease, asthma, COPD and idiopathic pulmonary fibrosis, depression and anxiety, frailty, brain aging and neurodegeneration, kidney disease, and now cancer.

The cancer paper is the one making the rounds. Zhu and Zhang, PNAS (2026), report hazard ratios of 0.31 for liver and bile-duct cancer, 0.48 for prostate, 0.59 for lung, 0.60 for rectum, 0.64 for breast, plus 0.05 SD longer telomeres — “approximately 2.2 fewer years of biological aging.” Look at the lung number and hold it still for a second. The claim is that whether your mother’s tea was sweetened in 1953 cut your lifetime lung cancer risk by 41%. Lung cancer in this cohort is overwhelmingly a smoking outcome; for an infant-nutrition hypothesis it is effectively a negative control, and it lit up as brightly as the outcomes that have a proposed mechanism. That is not corroboration. That is the falsification test failing and being reported as a finding. The contrast sits inside the same literature: a different group in AJCN (2026) ran the same instrument on early-onset cancer, got a more modest HR of 0.66 (0.53–0.81), did pre-specify negative controls — herpes zoster and cataract, both null — and found no effect on cancer-specific mortality. That is how you use a natural experiment.

There are two mechanical reasons this instrument keeps producing universal protection. First, there is no knife edge. Exposure is defined as the duration of the first 1,000 days spent under rationing, declining smoothly from full for a January 1951 birth to zero by mid-1956. The estimator is therefore a birth-date ramp across four and a half years, and it absorbs every secular trend in that window: smoking initiation, postwar recovery, maternal age, NHS access, vaccination schedules. Regression discontinuity earns its credibility from local comparison across a sharp threshold; this design compares 1951 to 1956. Second, left truncation. UK Biobank recruited in 2006–2010, so to be in the sample you had to be alive and willing then — someone born in 1951 had to survive to 55, someone born in 1956 only to 50. The more-exposed cohorts are thus more heavily pre-screened for people who did not die of cancer, heart disease or liver failure in their early fifties, and that screening is monotone in birth date, exactly like the treatment variable. It manufactures a dose-response curve out of nothing, simultaneously, for every fatal outcome. Which is what the literature reports. A tell of the researcher degrees of freedom involved: the same source population yields N = 64,761, 63,819, 63,793, 58,670 across four papers, and the respiratory one defines “unexposed” as born after July 1954 while the others use September 1953. Four rulebooks, one dataset. The cardiovascular paper has already required a formal correction.

The reframe I would argue for is that “we ran negative controls” is the wrong bar, because a cohort artifact sails straight past a badly chosen one — cataract isn’t fatal, so it won’t track a mortality-truncation gradient. The right diagnostic is coverage: what fraction of the phenome does your instrument protect against? A nutrient protects against some things. A birth-cohort dummy protects against all-cause mortality. Thirteen papers in, this instrument is converging on all-cause, and nobody has noticed, because each paper reports exactly one disease and the pattern only exists in the union. The fix is cheap and nobody is doing it: require any new paper on the 1953 cutoff to publish the full PheWAS — every outcome, one volcano plot — not just the one it came for. A symmetric cloud shifted protective is a cohort effect; sugar-plausible outcomes standing out above the null is a real result. That figure costs an afternoon and would settle the question. What I would bet on: the diabetes and hypertension findings largely survive replication in a design with true population sampling and no recruitment-age truncation, and the liver (0.31), prostate (0.48) and lung (0.59) cancer effects do not. I would put the lung result under 20% to replicate at anything near that magnitude — and it is the one that tells you how much to trust the other four.