What a food study can tell you

In a study published in 2013, two researchers took a cookbook, picked fifty common ingredients from random recipes, and searched the medical literature for studies linking each one to cancer. Forty of the fifty had them.¹ For most of what's in your kitchen, the research you'd need to be frightened of it, or reassured by it, already exists. This course is about how to tell which of those studies can bear the weight people put on them, and it starts here, with what each kind of food study is able to show.

That callout sits at the top of every lesson, word for word, because somebody may arrive at lesson 6 from a search engine and never see this one. Beat's number is its England line; its helplines page lists the rest of the UK. ANAD describes itself this way: "ANAD is a peer-run organization providing peer support. We do not provide therapy or anything medical."¹³ So it's somebody to talk to, not treatment.

What this course is, and what it leaves out

This course reads the evidence on food and health claim by claim: what's well established, what's argued over and by whom, and what would settle it. It won't tell you what to eat, and it leaves out:

• Diet plans and targets. No meal plans, calorie targets or macronutrient splits. An official number, such as a protein allowance, appears as a table to read, not a target.
• Vitamin and mineral biochemistry, except where a decision turns on it: supplements, deficiency, pregnancy.
• Clinical nutrition. Diabetes, kidney disease, eating disorders, food allergy, coeliac disease, children and pregnancy are named where they matter, with where to go, and not taught.
• Obesity medicine. The newer weight-loss drugs and surgery are named in lesson 4 and not evaluated.
• Sports nutrition beyond the protein evidence in lesson 5. Strength and Fitness, next on the Core, covers training.
• Food cost, culture and insecurity. They shape what anybody can eat, and this course didn't research them. Cooking Fundamentals, later this term, is where making food happens.
• Environmental and ethical arguments, such as meat and the climate, or animal welfare. They're value questions; lessons 8 and 9 name them as reasons people choose, without judging them.

Two free courses that cover more came up in this course's research: Wageningen's Nutrition and Health on edX, the more technical, and Stanford's Introduction to Food and Health on Coursera, the more practical.¹¹ This course copies one thing from Wageningen's 2018 syllabus: its first module covered "the different study designs that are used in nutrition research" before it taught a single nutrient.¹¹

Why food is harder to test than a pill

A company testing a blood pressure drug can give half its volunteers the drug and half an identical dummy pill, keep everyone from knowing which they got, count pills, and see what happens. Food breaks almost every part of that. The clearest statement of why comes from the field's defenders, not its critics. Satija, Yu, Willett and Hu, of Harvard's nutrition department, wrote a long answer to the critics in 2015, and this course read four of its sections in full.³ Four reasons come out of it, and a fifth sits under all of them.

You can't hide a diet. "Unlike classic drug trials, RCTs of dietary interventions typically cannot be blinded".³ (RCT means randomised controlled trial.) A woman told to cut her fat intake knows she's doing it.

People don't stay on diets. "Dietary interventions to promote weight loss routinely have dropout rates of 30–40% even after just 1 y of follow-up".³ A drug is a pill a day. A diet is three decisions a day, for years.

The outcomes are slow. The Physicians' Health Study below ran for about twelve years. The largest diet trial ever run, the Women's Health Initiative, reported a difference in deaths from breast cancer only in a follow-up at a median of 19.6 years, a result whose weight is argued over and which lesson 2 reads in full.¹⁷ Satija and colleagues say randomised trials with hard endpoints (outcomes such as heart attacks and deaths, rather than blood measures) are "usually not the most appropriate or feasible study design to answer nutritional epidemiologic questions regarding long-term effects of specific foods or nutrients (unless they can be packaged in a pill)."³ Hold on to that bracket; it comes back below.

Everything you eat is tangled up with everything else. John P. A. Ioannidis, the field's best-known critic, starts from the observation that in recent cohort meta-analyses "almost all foods revealed statistically significant associations with mortality risk", which he reads as a sign of accumulated bias rather than of real effects.² He argues that because nutrients are intercorrelated, one real cause can throw up many false associations, and that confounding cannot be fully handled. That's this course's summary; his own words on the second point: "no currently available cohort includes sufficient information to address confounding in nutritional associations."²

And under all four: you cannot change one thing. If you eat less of anything, you eat more of something else, or less in total. A pooled study of heart disease put it in one sentence: "Lower habitual intakes of SFAs, however, require substitution of other macronutrients to maintain energy balance."⁸ (SFAs are saturated fatty acids. Macronutrients are fat, carbohydrate and protein, and energy balance is calories eaten against calories used, the subject of lesson 3.) A drug trial adds a pill and leaves the rest of your life alone. A diet trial cannot. That gets its own section below.

Five designs, and what each can show

Sleep lesson 3 taught the split between experiments and observational studies, and why an association in a cohort cannot tell you which way the arrow points. This lesson adds the designs nutrition relies on.

This course's research file (its notes on every source it read, and how much of each) holds no nutrition case-control study, so that row is the course's own description, and the line about memory is its reasoning, not a finding.

The cohort

A cohort's strength is that it can watch what people really eat, over the decades diseases take. Its weakness is that the people who eat more vegetables are also, on average, different in other ways.

The defence makes a strong claim for cohorts, in its own words: "Well-conducted prospective cohort studies thus can be used to infer causality with a high degree of certainty when randomized trials of hard endpoints are impractical."³ The same paper concedes the limit: "the critical assumption of “no unmeasured or residual confounding” that is needed to infer causality cannot be empirically verified in observational epidemiology".³ Both sentences are from the same authors. Keep them together.

The controlled feeding study

At the other extreme sits the metabolic ward, where volunteers live in a research unit and eat only what they're given. Kevin Hall and colleagues at the US National Institutes of Health fed seventeen men a high-carbohydrate diet for four weeks, then a ketogenic diet (very low in carbohydrate, high in fat) with the same calories for four more. They measured energy use in metabolic chambers, sealed rooms that measure it, and with doubly labelled water, which lesson 2 explains.⁵ This course read the abstract.

The precision is the point. The ketogenic diet "coincided with increased EEchamber (57 ± 13 kcal/d, P = 0.0004)": a change in energy use of 57 calories a day, give or take 13, in the chamber. (EE is energy expenditure.) Doubly labelled water, averaged over the last two weeks of each diet, put it at 151, give or take 63: a bigger number, measured far less tightly. The authors conclude carefully: the diet "was not accompanied by increased body fat loss but was associated with relatively small increases in EE that were near the limits of detection with the use of state-of-the-art technology."⁵

Notice two things. This study wasn't randomised: everybody ate the same diets in the same order, so a ward buys control of the food, not of everything else. And it didn't count heart attacks; in eight weeks it couldn't have. A ward can answer a short question almost exactly and a long one not at all, because nobody lives on a ward for twenty years.

The randomised trial

A randomised trial is the only design here where chance, not the person, decides who gets what, and that's what lets it speak about cause. But it speaks about cause only for what was assigned. In a diet trial what's assigned is usually advice, so the result is the effect of being told to eat a certain way, taken as people actually took it. Ioannidis calls this "intention-to-eat" data, and wants more of it.²

Mendelian randomisation

Some people carry gene variants that shift what they eat or how they handle it. Because genes are shuffled at conception, comparing carriers with non-carriers is like a trial nature ran. A 2009 review by Schatzkin and colleagues names the lactase gene, LCT, as a stand-in for dairy intake, and puts the advantage this way: "Such a genetic proxy is measured with little error and usually is not confounded by nongenetic characteristics."⁷

The catch is in the same abstract. "Necessary assumptions are that the gene is independent of cancer, given the exposure, and also independent of potential confounders", and the sample sizes needed "are shown to be potentially daunting".⁷ Put plainly, the gene must reach the disease only through the food. When a gene has effects of its own on the disease, which geneticists call pleiotropy, the method breaks. A 2022 review says the method, "when robustly performed, is generally less prone to confounding, reverse causation and measurement error than conventional observational methods and has different sources of bias".⁶ Different, not none. Both reviews were read at abstract level. For most foods there's no good genetic stand-in, so few lessons here can use the design; lesson 10 uses it on alcohol.

Beta-carotene: the cohorts, then the trials

Beta-carotene is the orange pigment in carrots and many other fruits and vegetables. The first big trial summed up the observational evidence in its opening line: "Epidemiologic evidence indicates that diets high in carotenoid-rich fruits and vegetables, as well as high serum levels of vitamin E (alpha-tocopherol) and beta carotene, are associated with a reduced risk of lung cancer."¹⁴

The cohort numbers look strong even on blood levels, not just remembered diets. Ioannidis, in 2018: "The relative risk of death for the highest vs lowest group of beta carotene levels in serum or plasma was 0.69 (95% CI, 0.59-0.80)."² That's about a third lower risk of death for people at the top against people at the bottom, with an interval nowhere near 1.

Put the two trials' death results beside the cohort estimate on one scale.

The Physicians' Health Study is not on the chart because its abstract gives death counts (979 on beta-carotene, 968 on placebo, in arms of almost equal size) rather than a risk ratio for death.¹⁶ Worked out from those counts, which is this course's arithmetic, its ratio is about 1.01, close to the dashed line.

Why the arrow reversed

There are two explanations, each from people who know the evidence well. This course thinks both are true.

The critic's reading: the cohorts were confounded. A high blood level of beta-carotene marks a diet rich in fruit and vegetables and, plausibly, not smoking and other healthy habits. (That's the research file's reading, not something the cohorts measured.) Adjusting for what you measured cannot remove what you did not. Ioannidis, in the sentence after his 0.69: "Even when measurement error is mitigated with biochemical assays (as in this example), nutritional epidemiology remains intrinsically unreliable."²

The defender's reading: the trials asked a different question. The cohorts measured beta-carotene as it arrives, inside food, as part of a whole diet. The trials gave large doses of one isolated compound as a pill, often to smokers. That's the bracket from earlier, a nutrient packaged in a pill. Satija and colleagues' general version: "an equally, and sometimes more likely, possibility is that the RCT and observational studies are answering very different questions."³ And on supplements specifically: "High-dose vitamin and antioxidant trials mimic drug trials that examine the effect of isolated compounds, but these findings have been largely negative".³

Both readings agree the pills failed. They part over what the failure says about cohort research as a whole, and that can be tested with data, which is what the next section does.

Do trials and cohorts usually disagree?

Beta-carotene is dramatic, but a famous case can't tell you how often it happens. For that somebody has to line up many matched pairs of trial and cohort evidence, and in 2021 Schwingshackl and colleagues did.⁴ This course read the abstract and two passages of the full paper.

They found 97 diet-disease outcome pairs, each with a body of trial evidence and a body of cohort evidence matched as closely as the authors could on who was studied, what they got, what it was compared with and what was counted. Some matched well and some didn't, and that turns out to matter. For each pair with a yes-or-no outcome, such as a heart attack or a death, they divided the trial result by the cohort result. This is a ratio of risk ratios. If trials and cohorts agreed perfectly, it would be 1.

Reading 1.09

Take the authors' own example first: a trial risk ratio of 0.95 against a cohort risk ratio of 0.90 gives 0.95 ÷ 0.90, about 1.06. The trial found a little less benefit. But a trial finding 1.06 against a cohort finding 1.00, a little more harm, also gives 1.06. So they warn that "the ratio of risk ratios should not be interpreted as larger or smaller treatment effects in one type of study".⁴ A pooled 1.09 means the two bodies of evidence differ a little on average. It doesn't mean trials find smaller effects.

That's a partial vindication for the defence: on average, cohort evidence lands near trial evidence.

But read the next number. The prediction interval, 0.81 to 1.46, says where the ratio for a single new pair might fall. That's a different thing from the confidence interval, 1.04 to 1.14, which says where the average ratio probably lies. The prediction interval is much wider because the pairs disagreed with each other a good deal, which is what the I2 of 68% in the quotation measures. (You can skip the τ2.) The authors spell it out: "The prediction interval indicated that the difference could be much more substantial, in either direction."⁴ Their conclusion keeps the same balance: "important differences or potential bias in individual comparisons or studies cannot be excluded."⁴ This is the critic's point, and it comes from the same abstract. On average they agree; for any one question you cannot assume it.

Now you try one.

Instead of what?

Every Core course before this one has added a question to the institute's way of reading a claim. Memory added the sample, Focus and Deep Work the instrument, Note-Taking the setting, Sleep the clock, and Mental Fitness what the comparison group got. This course adds one more: instead of what?

A claim about a food or a nutrient is a claim about a swap, because you cannot change one thing. So "cut saturated fat" isn't one claim. It's several, and they can have different answers.

The cohort evidence shows it plainly. Li and colleagues, publishing in 2015, followed 84,628 women and 42,908 men for 24 to 30 years, asking about diet every four years, and modelled what happened when 5 percent of energy moved from saturated fat to something else.⁹ This course read the abstract.

Same nutrient removed; four answers, depending on what came in. An earlier pooled analysis of eleven cohorts, also read at abstract level, found no association for monounsaturated fat, where Li found 15% lower, so that row has two answers; lesson 6 sets them side by side.⁸

The World Health Organization's 2023 guideline, whose summary this course read in full, explains why some studies of lower saturated fat show no benefit: in them, "the nutrients replacing SFA may themselves increase the risk of disease and therefore may mask any benefit of reducing SFA intake. Consequently, choice of replacement nutrient is key to obtaining a health benefit from reducing SFA intake."¹⁰

The other side has a real objection to the table. The scientific document published alongside the US government's 2025-2030 Dietary Guidelines in January 2026, whose fats chapter this course read in full, says "the statistical constructs used in cohort substitution models infer hypothetical nutrient exchanges that did not actually occur in a person's diet", and that "higher linoleic acid intake and modeled substitution for saturated fat may partly reflect healthy user/adherer bias despite multivariable adjustment."¹² (Linoleic acid is an omega-6 polyunsaturated fat, the kind in the corn oil some of the old trials used. Healthy user bias means that people who follow health advice in one thing tend to follow it in others.) It is a fair methodological point: nobody in those cohorts was told to swap anything, and the swap is a model. Lesson 6 weighs the trial evidence against it, and lesson 9 has the 2026 guidelines themselves, as they stood in September 2026.

For this lesson the point is narrower. The mainstream reads the table and says the replacement decides the answer. The dissent says the table's swaps are hypothetical. Either way, you cannot evaluate "cut saturated fat" until somebody says what's replacing it.

How this differs from Mental Fitness's question

Mental Fitness lesson 2 asked of every effect: what did the comparison group get? That's about the arm of a trial (a waiting list, a placebo, another treatment), and it sets how big a number looks.

"Instead of what?" is about the plate, not the arm. It applies to a cohort, where nobody assigned anything, as much as to a trial. And in a diet trial the two questions come apart. A trial can compare a low-fat diet group with a control group who got leaflets; that's the comparison group. Inside the low-fat group, the fat that came off the plate was replaced by something; that's the swap. You'll often need to ask both.

What people get wrong

"Observational nutrition studies are worthless." On average, in 97 matched pairs, cohorts and trials came out close, and the big disagreements came mostly from pairs that asked different questions.⁴ Cohorts are the only design that can watch real diets for the decades diseases take. The critics' point survives in a narrower form: any single cohort finding can be confounded, and a given pair can disagree a lot.

"A randomised trial settles any food question." A trial settles the question it actually tested, as its participants actually carried it out. A diet trial usually cannot be blinded, loses people, and may not run long enough for the disease to show.³ Ioannidis, the critic, is candid about this too: "Large pragmatic trials for more complex diet patterns also may yield largely negative results."² He still wants them.

"A supplement trial tests the food." It tests the pill. Beta-carotene in a 20 or 30 mg capsule, given to smokers, is a different exposure from beta-carotene in carrots. In both trials that gave them to smokers, the pills raised lung cancer and deaths. They didn't test carrots.¹⁴¹⁵ (The US Preventive Services Task Force now recommends against beta-carotene supplements for preventing cancer or heart disease; lesson 10 reads that recommendation.)¹⁸

"A strong, consistent cohort result, measured in blood rather than memory, is close to proof." Beta-carotene had all of that: a risk ratio of 0.69 with an interval nowhere near 1, on a blood measure rather than a questionnaire.² Pills then raised deaths in two trials. The defenders concede the point in principle: the assumption of no unmeasured confounding "cannot be empirically verified in observational epidemiology".³ A narrow interval rules out chance, not confounding.

One more, because the critics' most famous argument is easy to misread. Ioannidis calculated that if the cohort meta-analyses showed lifelong causal effects, then "for a baseline life expectancy of 80 years, eating 12 hazelnuts daily (1 oz) would prolong life by 12 years (ie, 1 year per hazelnut)".² That sentence begins "Assuming the meta-analyzed evidence from cohort studies represents life span–long causal associations". It's a reductio, meant to show the assumption is absurd, not a claim about hazelnuts.

Practice

Connections

Back. Sleep lesson 3 taught experimental against observational evidence, Mental Fitness lesson 2 what the comparison group got, and Reading Well lesson 8 how to read a research paper, which is how every study here was read, most only at the abstract.

Forward. Lesson 2 turns to the instrument under all the cohorts, the food questionnaire, and what each side makes of its errors. It also reads the Women's Health Initiative and asks: its low-fat arm ate less fat instead of what?

Go deeper

• Satija, Yu, Willett and Hu, 2015, free at PubMed Central. Read here: the abstract and four sections. The defence of cohort research from Harvard's nutrition department, two of its senior figures among the authors, concessions included.
• Ioannidis, 2018, paywalled at JAMA. Read in full here. The critique in two pages, including what he thinks trials can and cannot do.
• Schwingshackl and colleagues, 2021, free at PubMed Central. Read here: the abstract and two passages of the full text. The best single answer to "do cohorts and trials agree?"
• Wade and colleagues, 2022, free at PubMed Central. Abstract only here. A long introduction to Mendelian randomisation in nutrition.

Sources

1. J. D. Schoenfeld and J. P. A. Ioannidis, "Is everything we eat associated with cancer? A systematic cookbook review", American Journal of Clinical Nutrition 97(1), 2013, pp. 127 to 134, doi 10.3945/ajcn.112.047142. Read: the abstract. Supports: fifty ingredients, forty with studies. The abstract does not name the cookbook.
2. J. P. A. Ioannidis, "The Challenge of Reforming Nutritional Epidemiologic Research", JAMA 320(10), 2018, pp. 969 to 970, doi 10.1001/jama.2018.11025. Read: the full text (2 pages), from a reprint of the publisher's PDF. Supports: every quotation from him. The argument about intercorrelated nutrients is the research file's summary of his piece, not a quotation.
3. A. Satija, E. Yu, W. C. Willett and F. B. Hu, "Understanding nutritional epidemiology and its role in policy", Advances in Nutrition 6(1), 2015, pp. 5 to 18, doi 10.3945/an.114.007492. Read: the abstract and, in full, the sections "Introduction", "Can We Reliably Measure Dietary Intakes...", "What Is the Role of Nutritional Epidemiology in Inferring Causality?" and "Is the Drug Trial Paradigm Relevant..."; the rest skimmed. Supports: every quotation from the defenders.
4. L. Schwingshackl and colleagues, "Evaluating agreement between bodies of evidence from randomised controlled trials and cohort studies in nutrition research: meta-epidemiological study", BMJ 374, 2021, n1864, doi 10.1136/bmj.n1864. Read: the abstract, and from the full text the Methods paragraph on how pairs were compared and the Results' first paragraph. Supports: 97 pairs, 1.09, its intervals, the explanation and the conclusion (abstract); the 71 binary pairs, the 0.95 against 0.90 example and the warning (full text).
5. K. D. Hall and colleagues, "Energy expenditure and body composition changes after an isocaloric ketogenic diet in overweight and obese men", American Journal of Clinical Nutrition 104(2), 2016, pp. 324 to 333, doi 10.3945/ajcn.116.133561. Read: the abstract. Supports: the design, 57 and 151 kcal a day, and the conclusion.
6. K. H. Wade and colleagues, "Applying Mendelian randomization to appraise causality in relationships between nutrition and cancer", Cancer Causes and Control 33(5), 2022, pp. 631 to 652, doi 10.1007/s10552-022-01562-1. Read: the abstract.
7. A. Schatzkin and colleagues, "Mendelian randomization: how it can--and cannot--help confirm causal relations between nutrition and cancer", Cancer Prevention Research 2(2), 2009, pp. 104 to 113, doi 10.1158/1940-6207.capr-08-0070. Read: the abstract. Supports: the lactase gene, the advantage, the assumptions and the sample sizes.
8. M. U. Jakobsen and colleagues, "Major types of dietary fat and risk of coronary heart disease: a pooled analysis of 11 cohort studies", American Journal of Clinical Nutrition 89(5), 2009, pp. 1425 to 1432, doi 10.3945/ajcn.2008.27124. Read: the abstract. Supports: the substitution sentence and no association for monounsaturated fat.
9. Y. Li and colleagues, "Saturated Fats Compared With Unsaturated Fats and Sources of Carbohydrates in Relation to Risk of Coronary Heart Disease: A Prospective Cohort Study", Journal of the American College of Cardiology 66(14), 2015, pp. 1538 to 1548, doi 10.1016/j.jacc.2015.07.055. Read: the abstract. Supports: the cohorts and every row of the table.
10. World Health Organization, Saturated fatty acid and trans-fatty acid intake for adults and children: WHO guideline summary, 2023, landing page. Read: the guideline summary in full; the 134-page full guideline was not opened.
11. Wageningen University, Nutrition and Health: Macronutrients and Overnutrition (edX, NUTR101x), read: the official syllabus PDF, dated 12/13/2018, for the 1T2018 run; the current edX page may differ. Stanford Online, Introduction to Food and Health (Coursera), course page read as of 24 September 2026.
12. US Department of Health and Human Services and US Department of Agriculture, The Scientific Foundation for the Dietary Guidelines for Americans, 2025–2030, January 2026, PDF. Read: the front matter and Chapter 5, "Fats and Oils", in full.
13. ANAD, "Get help", anad.org, and Beat, helplines page, both fetched on 24 September 2026. Supports: the callout's numbers and ANAD's description. Hours change, so the lesson links to the pages rather than printing them.
14. The Alpha-Tocopherol, Beta Carotene Cancer Prevention Study Group, "The effect of vitamin E and beta carotene on the incidence of lung cancer and other cancers in male smokers", New England Journal of Medicine 330(15), 1994, pp. 1029 to 1035, doi 10.1056/NEJM199404143301501. Read: the abstract.
15. G. S. Omenn and colleagues, "Effects of a combination of beta carotene and vitamin A on lung cancer and cardiovascular disease", New England Journal of Medicine 334(18), 1996, pp. 1150 to 1155, doi 10.1056/NEJM199605023341802. Read: the abstract.
16. C. H. Hennekens and colleagues, "Lack of effect of long-term supplementation with beta carotene on the incidence of malignant neoplasms and cardiovascular disease", New England Journal of Medicine 334(18), 1996, pp. 1145 to 1149, doi 10.1056/NEJM199605023341801. Read: the abstract. Supports: the trial, the death counts, the conclusion and its authors.
17. R. T. Chlebowski and colleagues, "Dietary Modification and Breast Cancer Mortality: Long-Term Follow-Up of the Women's Health Initiative Randomized Trial", Journal of Clinical Oncology 38(13), 2020, pp. 1419 to 1428, doi 10.1200/jco.19.00435. Read: the abstract. Supports: the 19.6-year median follow-up. Lesson 2 reads this trial properly.
18. US Preventive Services Task Force, "Vitamin, Mineral, and Multivitamin Supplementation to Prevent Cardiovascular Disease and Cancer: US Preventive Services Task Force Recommendation Statement", JAMA 327(23), 2022, pp. 2326 to 2333, doi 10.1001/jama.2022.8970. Read: the abstract in full, and the recommendation web page. Supports: the recommendation against beta-carotene supplements.