The Outcome Evidence: Associations, Confounding, and Dose

The hardest questions about ultra-processed food — does it cause disease, and how much is too much — cannot be answered by randomized trials, because no ethics board will assign people a decade of packaged food. What we have instead are enormous observational cohorts: hundreds of thousands of people, followed for years, their diets classified and their outcomes tracked. This page lays out what those cohorts found for weight, cardiometabolic disease, cancer, and mortality — and what that evidence can and cannot tell you.

What the Cohorts Found

The evidence base rests on a handful of very large prospective studies. The French NutriNet-Santé cohort — over 100,000 adults submitting repeated 24-hour dietary records, each food classified by NOVA processing level — produced the headline analyses, and the American Nurses' Health Study and Health Professionals Follow-up Study contributed decades of follow-up. Four outcome families dominate the literature: body weight, cardiometabolic disease, cancer, and mortality. The table below summarizes the strongest single result in each, expressed the way the studies report it: hazard ratio per additional 10% of the diet (by weight) from ultra-processed foods.

One feature deserves emphasis before the caveats arrive: the cancer and cardiovascular associations held after adjustment for nutritional quality — saturated fat, sodium, sugar, fiber, and Western dietary pattern — in both the 2018 cancer analysis (Fiolet et al., BMJ) and the 2019 cardiovascular analysis (Srour et al., BMJ). Whatever the linkage is, the models say it is not entirely the nutrient profile. That claim, and its limits, is where the rest of this page goes.

Why "Association" Is the Precise Word

A hazard ratio of 1.15 for type 2 diabetes means that, within the modeled population, each additional 10% of the diet coming from ultra-processed foods tracked with 15% higher incident-disease rates over follow-up — after the researchers adjusted for the confounders they measured. It does not mean eating those foods raised any individual's risk by 15%, and it does not mean the foods themselves did anything. The diabetes analysis shows the discipline involved: Srour and colleagues adjusted for weight change during follow-up and the association persisted (HR 1.13, 1.01–1.27) — not simply a weight-gain story retold — and grams-per-day intake predicted risk independently of the unprocessed food eaten alongside it.

The absolute framing matters just as much. In the cardiovascular cohort, the highest-intake quarter experienced 277 cardiovascular events per 100,000 person-years against 242 in the lowest — 35 extra cases per 100,000 person-years, a genuinely elevated rate and a modest one at the same time. Both numbers are true.

The Confounding Problem, Examined Honestly

Who eats the most ultra-processed food? The NutriNet mortality analysis (Schnabel et al., JAMA Internal Medicine, 2019) answers with its own baseline table — the confounding problem made visible:

• 📉 Higher intake tracked with disadvantage — younger age, lower income, less education, and living alone all rose with ultra-processed share; each of these independently predicts the outcomes studied.
• 🚭 Lifestyle clusters — higher BMI and lower physical activity sat in the same high-intake rows; smoking is the classic phantom in diet cohorts, and pack-years adjustment barely moved the US mortality estimate.
• 🔄 Reverse causation runs too — illness, fatigue, and tight budgets push people toward convenient food, so early disease can drive the diet rather than the reverse; better studies exclude early deaths and lag the diet data to blunt this.
• 🧩 Residual confounding never hits zero — statistical models adjust for what was measured, imperfectly; unmeasured or coarsely measured factors remain. Every serious author in this field states this limitation. So does this page.

This is precisely why the controlled-feeding trial matters as the causal complement: randomization balances the confounders observational models can only approximate. But that trial measured intake over four weeks in 20 adults — it established that ultra-processed menus drive overeating, not that decades of them cause disease. The two evidence types meet in the middle: plausibility from the trial, population-level association from the cohorts, and a gap between them that honesty requires naming.

Dose, and Where the Signal Concentrates

Dose-response is the strongest card observational evidence can play, and here it is strong: across the pooled analyses, risk rises roughly linearly with each 10% increment, with no threshold below which the association vanishes — a gradient, not a cliff. What the dose story does not tell you is which foods carry it. The US mortality analysis (Fang et al., BMJ, 2024) followed 74,563 women and 39,501 health professionals for over three decades and found a much smaller headline — 4% higher all-cause mortality in the highest quarter (~7.4 servings/day) versus the lowest — driven mainly by causes other than cancer or cardiovascular disease, with cancer and cardiovascular associations too weak to rely on. Its subgroup analysis was the interesting part: meat, poultry, and seafood-based ready-to-eat products showed the most consistent associations, while grain-based subgroups showed little to none.

The 2024 umbrella review (Lane et al., BMJ) formalized the hierarchy: pooling 45 meta-analyses covering 9.9 million participants, it graded type 2 diabetes as convincing-class evidence (dose-response risk ratio 1.12 per 10%), with cardiovascular mortality and obesity one tier down — and most other outcomes suggestive or weak. Under GRADE criteria, 22 pooled analyses were low quality and 19 very low. Honest summary: the direction is consistent, the strongest associations are cardiometabolic, and the quality is limited by design. Which subcategories matter most — sugary drinks, processed meat, texture-engineered snacks — is explored in the mechanisms page, alongside the nutrient overlap owned by fructose's special case, fats and carbs, and fiber.

Reading the Numbers Like an Epidemiologist

• 📐 Per-10% compounds — moving from 30% to 50% of the diet as ultra-processed food stacks two increments; the cohorts price each one, and few real diets sit at a single step.
• 🎚️ Exposure measurement is soft — food-frequency questionnaires misclassify intake, an error that usually dilutes true associations; the French repeated-records design is stronger, and it found the larger effects.
• 🗺️ Populations differ — NutriNet volunteers are health-conscious web-recruits with low average intake (~14% of food weight); US cohorts start higher. Estimates are population averages, not personal forecasts.

Questions, Answered Briefly

• 🧪 "Is any of this randomized?" For disease endpoints, no — the trial evidence covers short-term intake and weight, which the feeding-trial page handles. Everything on this page is observational.
• 🥗 "Is it the nutrients or the processing?" Adjustment for nutrient profile attenuates the associations without erasing them (Fiolet et al., BMJ, 2018; Srour et al., BMJ, 2019) — suggesting processing-level factors contribute, unresolved.
• 📏 "How much is too much?" The cohorts show a continuous gradient with no identified threshold; they price each 10% increment rather than drawing a line. For what to swap first, see the parent guide's ultra-processed foods topic.

Related Topics