Nutrition advice appears to reverse itself constantly, which is usually blamed on scientists being unreliable. The actual explanation is that the research is extraordinarily difficult to do well.
You cannot run the ideal study
The gold standard would randomise people to different diets and follow them for decades with perfect compliance.
Which is impossible. People will not eat assigned diets for decades, controlled feeding is enormously expensive, and blinding is impractical for most dietary interventions.
So most nutrition evidence comes from observational studies, which follow what people eat and what happens to them.
The measurement problem
Observational studies rely on people reporting what they ate, generally through questionnaires.
Self-reported dietary intake is known to be inaccurate. People underreport, particularly foods they consider undesirable, and misremember quantities.
Validation studies comparing reported intake against objective measures find discrepancies large enough to matter.
Which means the exposure variable in most nutrition research is measured with substantial error, and error in the exposure weakens and distorts any relationship found.
Confounding
The larger problem.
People who eat more of one thing differ systematically from those who do not — in income, education, exercise, smoking, healthcare access and a hundred other things.
Studies adjust statistically for known confounders, which handles what is measured and cannot handle what is not.
The healthy user effect is the standard example. People who follow one piece of health advice tend to follow others, which makes any single behaviour look more beneficial than it is.
This has produced findings that were later contradicted by trials, most famously in vitamin supplementation.
Substitution is implicit
A subtle point that changes interpretation considerably.
Eating less of something means eating more of something else, and the comparison is always against whatever replaced it.
Which means a finding that reducing one nutrient improves outcomes depends entirely on what people ate instead, and studies frequently do not specify.
This explains several long-running controversies where both sides were partly right about different substitutions.
Effect sizes are small
Relative risk figures in nutrition research are typically modest.
Which matters because small effects are exactly where confounding and measurement error are most likely to produce spurious results.
A large relative risk is hard to explain away. A small one is not.
Headlines convert small relative risks into alarming statements by omitting the absolute risk, which is frequently very low.
Funding
Industry funding of nutrition research is widespread, and analyses have found association between funding source and reported conclusions.
Which does not mean funded research is wrong and does mean funding is relevant context that is often buried.
What holds up
A small number of findings have survived across study designs and decades.
Eating enough vegetables, fruit, legumes and whole grains is associated with better outcomes consistently.
Very high intakes of processed meat and of added sugar are associated with worse outcomes consistently.
Adequate protein and micronutrients matter, and deficiency causes identifiable disease.
Total energy balance determines weight, though the mechanisms influencing intake are considerably more complex than willpower.
Beyond that, most specific claims are less established than their confidence suggests.
How to read a nutrition headline
Ask whether it was a trial or an observational study.
Ask what the absolute risk change was.
Ask what the comparison food was.
And note that a single study rarely changes anything, which is why the reversals people notice are generally media coverage reversing rather than the evidence.
Anyone with specific dietary requirements or a medical condition should be talking to a registered dietitian rather than reading general commentary.
Mendelian randomisation
A method that partially addresses the confounding problem and is worth knowing about.
It uses genetic variants associated with a particular exposure as a proxy for that exposure.
Because genetic variants are allocated at conception and are not affected by lifestyle, they are less subject to confounding than measured behaviour.
Which has produced findings that contradicted observational results in several areas, and has supported them in others.
The method has its own assumptions and limitations, and it has become one of the more useful tools where trials are impossible.
The processing question
Currently the most active area and the most contested.
Classification systems grouping foods by degree of industrial processing have produced consistent associations between highly processed foods and poor outcomes.
The criticism is that the classification is imprecise, groups very different products together, and may be capturing overall diet quality rather than processing itself.
A controlled feeding trial found higher energy intake on a processed diet than on a matched unprocessed one, which is the strongest experimental evidence available and rests on a single small study.
Individual variation
An area of active research that complicates general recommendations.
Studies measuring blood glucose responses to identical meals have found substantial variation between individuals, attributed partly to gut microbiome composition.
Which has generated commercial personalised nutrition services, ahead of evidence that acting on the measurements improves health outcomes.
The underlying observation of variation is solid. The claim that it can currently be used to produce better individual advice is not.