A preprint says ultra-processed food raises frailty risk 43%. Here is how to read a preprint.

A paper posted today on medRxiv pools ten observational studies of ultra-processed food and frailty in older adults — 105,000 participants across the United States, Spain, Italy, Australia, Singapore, Iran and China — and reports that higher intake was associated with 43% higher odds of frailty. The direction fits a large literature, including the American Heart Association advisory we covered in August, and the headline will travel. Before it does, this is a good specimen for a skill every reader of health news now needs: how to read a preprint. A preprint is a paper that has been posted before peer review. Nobody has yet checked it who did not write it.

Four things to look at, in order. First, the uncertainty: the 43% comes with a range of 2% to roughly 100%, which means the data are consistent with almost no effect and with a doubling. Second, consistency: the pooled studies disagreed with one another to an extreme degree — the heterogeneity statistic was close to 98%, which tells you the single number is averaging studies that measured different things in different ways. Third, publication bias: the authors' own test suggests small studies with positive results are over-represented, and removing studies one at a time did not consistently keep the result significant. Fourth, who did it: two authors, no funding, no declared conflicts, which is honest and also a reason the work is small. None of this makes the paper wrong. It makes the 43% a hypothesis, which is what a preprint is.

Loree’s take

I will see this number in my inbox by the weekend. What I will say is that the ultra-processed-food literature does not need this paper — the signal across large cohorts is consistent — and that this particular estimate is the weakest kind: wide, heterogeneous and unreviewed. Frailty in older adults is real and food is part of it. If a reader eats more real food because of a shaky 43%, no harm done. If a reader learns to recognize a shaky number, better.

Ben’s take

Preprints exist so scientists can see work early. They are now harvested by marketers who do not care whether the review ever happens, because a URL with a DOI looks like proof in a product page. The defense is the four questions above. They take two minutes and work on every preprint, including the ones that agree with us. Especially those.

Why it matters to a recovering optimizer. The longevity trap runs on numbers that arrived before the checking did. A preprint is that process made visible, with a label on it. Learn to read the label and you will be harder to sell to — by the industry, and by anyone whose conclusion you already like, which includes this site.