Ultra-processed food: the trial that moved it past correlation

The evidence problem comes first

When the authors of the largest synthesis in this field went looking for pooled analyses of randomised trials, they found none, and narrowed their scope to observational epidemiology.¹ That is the honest starting point.

People who take most of their energy from ultra-processed products are not a random sample: they tend to be poorer, to smoke more, to move less and to eat more of everything, and each of those is also a cause of the diseases under study. Adjustment only works on variables that were measured, measured well and modelled correctly. The exposure is measured badly too — almost every cohort estimates it from food frequency questionnaires and dietary recalls, instruments built for nutrients rather than processing, which the reviewers concede introduces an inevitable measurement bias.¹

None of which makes the picture uninformative. A review of 43 studies found 37 reporting an association with at least one adverse outcome and none reporting a benefit.⁴ Consistency like that is worth something; it is not a demonstration.

What a metabolic ward can do that a cohort cannot

In 2019 Kevin Hall's group at the US National Institutes of Health ran the trial the field had been missing.² Twenty weight-stable adults, mean age 31.2 ± 1.6 years and mean BMI 27 ± 1.5, spent 28 days on the metabolic ward at the NIH Clinical Center, randomised to two weeks of an ultra-processed diet then two weeks of an unprocessed one, or the reverse.

The point was the matching. Two seven-day rotating menus were built to be equivalent in presented calories, energy density, macronutrients, sugar, sodium and fibre, and to differ chiefly in the share of energy from NOVA group 4 foods: 83.5% against 0%. Food was presented at twice each participant's estimated requirement, to be eaten ad libitum.

Participants ate 508 ± 106 kcal per day more when the food was ultra-processed (p = 0.0001) — extra carbohydrate (280 ± 54 kcal/day) and extra fat (230 ± 53), with protein unchanged (−2 ± 12, p = 0.85). Weight tracked intake closely (r = 0.8): they gained 0.9 ± 0.3 kg on the ultra-processed arm and lost 0.9 ± 0.3 kg on the unprocessed one. They also ate the ultra-processed meals 17 ± 1 kcal per minute faster (p < 0.0001), though both menus were rated equally pleasant and familiar. The paper reports standard errors, not confidence intervals, throughout.

The limitations are not small

Twenty people, two weeks an arm, one metabolic ward. There was no run-in and no washout, and the authors decline to claim more than they tested: Our study was not designed to identify the cause of the observed differences in energy intake.²

The structural limitation is visible in Table 1. Total energy density was matched almost exactly, 1.024 against 1.028 kcal per gram; the energy density of the non-beverage food was not, at 1.957 against 1.057, because the matching ran partly through drinks. You cannot hold everything constant while changing processing, because processing is not an ingredient.

Had the menus differed in sugar, fat and sodium as much as real diets do, the authors note, we may have observed larger differences in energy intake.² A correction published later that year fixed a documentation error in one participant's meal orders without touching the primary outcome.⁵

A second trial, outside the ward

Nobody lives on a metabolic ward. In 2025 a British group ran the complement.³ Fifty-five adults in England, all with a BMI of 25 to 40 and habitually taking at least half their calories from ultra-processed food, were given two eight-week ad libitum diets in random order with a four-week washout, every meal and drink delivered to their homes; 50 formed the intention-to-treat sample. Both diets complied with the UK's Eatwell Guide: guideline-compliant ultra-processed food against guideline-compliant minimally processed food.

Both produced weight loss; the minimally processed diet produced roughly twice as much. −2.06% of body weight (95% CI −2.99 to −1.13) against −1.05% (−1.98 to −0.13), a within-participant difference of −1.01% (−1.87 to −0.14; p = 0.024), and 0.98 kg more fat mass lost on it (standard error 0.32, p = 0.004).

For an 80 kg adult that is under a kilogram, and the caveats deserve equal billing: adherence was self-reported from diaries returned by only 32 and 35 of the 50 participants, and carryover cannot be excluded. The ultra-processed diet was again the more energy dense, 1.60 against 1.25 kcal per gram — the same asymmetry the ward produced.

NOVA, and the fair complaint about it

Both trials, and every cohort in the umbrella review, classify food with NOVA, which sorts foods into four groups according to the extent and purpose of the industrial processing they undergo.⁶ Group 4 is defined as formulations of ingredients, mostly of exclusive industrial use, that result from a series of industrial processes.

The standing objection is that this is a wide net, and the category does not behave as one thing. In the type 2 diabetes meta-analysis within the umbrella review, overall intake predicted higher risk while several subcategories ran the other way: ultra-processed cereals, wholegrain breads, packaged snacks and yoghurt were inversely associated.¹

Nor is it easy to apply. Asked to sort 120 marketed products and 111 generic food items into NOVA groups, 159 and 177 French food specialists reached a Fleiss' κ of 0.32 and 0.34, with only four items across both lists classified identically by everyone. They concluded that current NOVA criteria do not allow for robust and functional food assignments.⁷ Other assessments report stronger agreement, and the umbrella review cites them.¹ The dispute is open.

What the cohorts add, and what they cannot

The umbrella review pooled 45 analyses of cohort, case-control and cross-sectional studies covering 9,888,373 people, and found direct associations with 32 of the 45 health parameters examined.¹

The grades matter as much as the numbers. Applying GRADE, they rated 22 of the 45 analyses low quality and 19 very low; only four reached moderate. A risk ratio of 1.50 with a tight interval, sitting on very low quality evidence, is a precise summary of imprecise inputs.

What the cohorts add is scale, duration and hard endpoints no eight-week feeding trial can reach. What they cannot add is causal identification, and their authors say so: residual confounding is perhaps most pertinent among the limitations of observational work, though they argue consistency across analyses makes it unlikely to explain everything. Trials of long-term exposure measured against cardiovascular disease or cancer, they add, will not be possible, for obvious ethical reasons.¹

The trials show that, with nutrients held roughly level, replacing ultra-processed food with less processed food reduces how much people eat and weigh over weeks. The cohorts show the same exposure tracking disease and death over decades, with confounding unexcluded. Together they are a good deal more than a correlation.

All of this depends on seeing what you actually eat. Kettle logs food macros — protein, carbohydrates, fat and energy — and water alongside your training on iPhone, with data kept on the phone.