Meta-analysis in nutrition

Evidence synthesis in nutrition

Nutrition headlines often come from meta-analyses, and the field has been the source of debate about the strength of evidence. Questions such as whether red meat raises cancer risk, whether salt reduction lowers blood pressure and mortality, or whether a dietary pattern protects against heart disease are answered by combining large observational cohorts with smaller trials. The data are different in kind: cohorts show associations over decades but cannot control all confounders, while trials of diet can show causal effects on biomarkers over weeks or months but rarely run long enough to measure disease.

The field has a history of consistent findings overturned. Observational studies of antioxidant vitamins suggested protection against cardiovascular disease and cancer, but large trials of supplements found no benefit and, in some cases, harm. This is a standard example of why reviews should state the type of evidence and why certainty ratings matter. Methods for rating certainty in nutrition, including updates to GRADE and new tools for nutritional exposures, continue to develop.

The general framework is in meta-analysis and systematic review, and the notes on observational data are in MOOSE reporting.

Dose-response and intake categories

Nutritional epidemiology often reports risk by categories of intake. A meta-analysis of the highest versus lowest category is simple but loses information, because category boundaries differ between studies. Dose-response meta-analysis, using the method of Greenland and Longnecker and later one-stage models, estimates the change in risk per unit of intake, such as per 100 grams per day, and allows non-linear curves with splines. A reported relative risk of 1.08 per 10 grams per day implies, under a log-linear model, a relative risk of 1.08^(30/10) = 1.260 for 30 grams per day. The numbers are invented. Extrapolating beyond the range of intakes in the studies is not justified, and the plotted curve should show the range of data and the uncertainty.

Units must be harmonized across studies: servings differ in size between countries, and the nutrient content of a food varies. Reviews should state the conversion. Reference categories also matter, since a non-consumer group may differ from the lowest consumers in other ways. See dose-response meta-analysis for the methods.

Measurement error and confounding

Food frequency questionnaires and diet recalls are imprecise, and errors are correlated with factors such as body mass index and education. Random error biases associations toward the null, and systematic error can bias in either direction. Calibration studies using recovery biomarkers, such as urinary nitrogen for protein and doubly labeled water for energy, show that self-reported energy intake is underestimated, especially in people with obesity. Reviews should note the instrument, whether it was validated and whether the estimate was corrected for measurement error. Diet is also correlated with other behaviors: people who eat more vegetables tend to smoke less and exercise more. Studies adjust for these factors, but residual confounding remains. Reviews should extract the adjustment set and compare estimates with minimal and full adjustment, and may use negative controls or sibling comparisons if available.

Repeated dietary measures during follow-up capture changes in intake. Studies that use only baseline diet misclassify people whose diet changes. Substitution analysis, which estimates the effect of replacing one food by another, is more informative than adding a food to the diet while leaving everything else unchanged, since the effect of a food depends on what it replaces.

Trials of diet, supplements and biomarkers

Randomized trials of diet range from short feeding studies with controlled meals to long-term behavioral interventions. Long-term trials suffer from poor adherence: participants in the intervention and control groups may eat similarly by the end, which dilutes the contrast. Blinding is difficult for foods, easier for supplements and capsules. Many trials measure intermediate outcomes such as LDL cholesterol, blood pressure, glucose and weight, which are used as surrogates. The link between a biomarker and disease is not always straightforward, and an effect on a biomarker does not prove an effect on disease. The review should state the strength of the surrogate relationship.

Supplement trials use fixed doses, which can be tested in dose-response models, and baseline nutrient status modifies the effect: giving vitamin D to those deficient differs from giving it to those replete. Reviews should extract baseline status and analyze by subgroup. Conflicts of interest are important in nutrition, where industry funding is common and has been associated with favorable conclusions. Funding should be recorded.

Dietary patterns and network comparisons

Studies of Mediterranean, low-carbohydrate, low-fat, plant-based and other dietary patterns are numerous, and network meta-analysis has been used to compare them for weight loss and cardiometabolic risk. Defining a pattern is a challenge: the same name covers different diets in different trials, and adherence differs. Reviews should extract the macronutrient composition actually achieved and not rely on the label. Short-term weight loss differences between diets tend to shrink at one or two years, and the review should present results at stated time points. Because trials of diet are heterogeneous, the certainty of evidence is often low, and rankings should be presented with intervals and cautions.

Certainty of evidence in nutrition

GRADE starts evidence from randomized trials as high and from observational studies as low, and the latter can be upgraded for large effects, dose-response gradients and plausible residual confounding that would reduce the effect. Many nutrition questions depend on observational data, so certainty is often low or very low, and this has led to debate about how to present findings to the public. Newer tools, such as ROBINS-E for exposures, and the NutriGrade scoring system, aim to adapt rating to nutritional epidemiology. A review should explain its choice and show the reasoning for each rating. Statements such as "evidence suggests" should be tied to the rating, so readers can see the difference between a well-supported effect and a weak association.

Early life, pregnancy, older adults and global nutrition

Nutrition needs and risks change across the life course. Studies of infant feeding, such as breastfeeding and complementary foods, are mostly observational and strongly confounded by maternal education and income, and the few trials, such as cluster trials of breastfeeding promotion, give estimates for the promotion more than for the feeding itself. Nutrition in pregnancy is studied through supplements such as folic acid, iron and calcium and through dietary patterns, with outcomes in mother and baby that need long follow-up. In older adults, the questions include protein intake, vitamin D and malnutrition, where weight loss and illness both change diet and risk, which creates reverse causation. Reviews should record the age group and consider separate analyses.

In low- and middle-income countries, trials of micronutrient supplementation, food fortification, cash transfers and school feeding address undernutrition, stunting and anemia. Effects depend on the baseline prevalence of deficiency, the presence of infection and the delivery system. Cluster-randomized designs are common, and the review must handle clustering. Indicators like height-for-age z-scores depend on the growth reference, so the reference should be recorded. A pooled effect from settings with different baseline deficiency should be interpreted with attention to context.

Communicating findings carefully

Nutrition findings travel quickly into the news and into advice, so the way a review reports results matters. A relative risk of 1.2 for a food, from observational studies with low certainty, should not be reported as if the food causes disease. The absolute risk is often small, and the reference is rarely a non-consumer. A good report says what kind of evidence was used, how large the absolute difference is, how certain the evidence is, and what could change the conclusion. It also says when studies disagree and what the possible reasons are. When findings relate to a population recommendation, the review should avoid making recommendations itself, because guideline development involves values, costs and feasibility and has separate methods.

Readers of reviews with conflicts of interest should be able to see who funded the review and the underlying studies. Reviews funded by food companies have been found to reach more favorable conclusions in some analyses, and transparency about funding and the role of funders is part of the report.

Common pitfalls we look for

• Using highest versus lowest comparisons when dose-response data are available.
• Extrapolating curves beyond the range of observed intakes.
• Ignoring what a food replaces in the diet.
• Treating biomarker changes as proof of disease prevention.
• Mixing deficient and replete populations in supplement trials.
• Overlooking industry funding.

Planning and reporting

The protocol states the exposure and how it will be harmonized, the comparator, the outcomes, the designs included and how they will be analyzed (observational and trial evidence separately), and the plan for dose-response models. Searches cover MEDLINE, Embase, CENTRAL, CAB Abstracts and registries. Risk of bias uses RoB 2, ROBINS-I or ROBINS-E; certainty is rated with GRADE. Reporting follows PRISMA 2020 and MOOSE for observational studies, with a protocol in PROSPERO where the review is eligible.

How we support research projects in this area

Boundaries of this service

A review of nutrition studies describes associations and effects in groups. It does not provide dietary advice, and it does not tell a person what to eat, which depends on their health, preferences, culture and the advice of a dietitian or doctor. We do not give dietary plans or interpret an individual's laboratory results. Anyone with a medical condition or concerns about their diet should speak to a qualified professional.

Frequently asked questions