How good are the food questionnaires behind most diet headlines? Across 43 studies, a middling match for calories and a weak one for salt

Most of what you read about diet and long-term health starts with a form. It lists a hundred or more foods and asks how often you ate each one over the past months or year: never, once a month, twice a week, every day. Researchers call it a food frequency questionnaire. It is cheap and quick, and it can be sent to a hundred thousand people, which is why the big studies linking foods to heart disease, diabetes or cancer lean on it.

The obvious worry is whether anyone can answer it well. How often did you eat broccoli last year? So each questionnaire is supposed to be tested before use, by giving it to a group of people and comparing the result with a more detailed method. A review published in Food Science & Nutrition on 30 September 2026, from researchers in Catanzaro and Rome, pooled those tests.

They searched three databases for studies from January 2000 to February 2026 and started with 5,247 records. After screening they kept 43 studies with 8,467 healthy adults, all from Europe, North America, Australia or New Zealand. Each study compared a whole-diet questionnaire with one of two yardsticks: interviews about everything eaten in the previous 24 hours (25 studies), or a food diary kept as the person ate (18 studies). The typical questionnaire had 128 items, though they ranged from 16 to 322, and the typical study had 96 people.

The main result is a set of correlations, numbers between 0 and 1 that say how closely two measurements line up. For calories, the questionnaire matched the 24-hour interviews at 0.45 and the food diaries at 0.50. For protein the figures were 0.41 and 0.47, for fat 0.44 and 0.48. Fibre did best of the main nutrients, at about 0.6 against both yardsticks, and carbohydrate came in just above 0.5. The authors call this moderate-to-good.

It helps to know what 0.45 looks like. By our arithmetic, a correlation of that size means the questionnaire accounts for about a fifth of the differences between people on the comparison method. That is enough to do what these forms are mostly used for: splitting a large group into those who eat a lot of something and those who eat little. It is not enough to tell you how many calories or grams any one person ate, and the review says as much. Its conclusion is that the questionnaires are for ranking people.

Vitamins and minerals were less even. Riboflavin and magnesium did well, at 0.54 to 0.63. Sodium, which is what salt is counted as, managed 0.38 against the 24-hour interviews, and 0.17 once the figures were adjusted for how much people ate overall. That makes sense: nobody remembers how much salt went into the pan. Iodine was the weakest at 0.10 against the interviews, with a margin of error that includes zero, though only two studies measured it and a diary comparison gave a very different answer. The authors advise caution about using a general questionnaire to estimate any single vitamin or mineral.

The questionnaires also tended to give bigger numbers. Set against the 24-hour interviews, they came out higher for protein, fat, fibre, vitamin C, vitamin E, magnesium, potassium and sodium, among others. Against food diaries the gaps were smaller and mostly not clear, apart from protein, vitamin B6 and iron. Calories themselves did not differ clearly on either comparison.

Now the catch. Every yardstick here is also a person reporting their own food. Interviews, diaries and questionnaires share some of the same blind spots, such as forgotten snacks and optimistic portion sizes, and when two methods share an error they agree with each other more than either agrees with the truth. The review left out studies that checked questionnaires against blood or urine markers, which do not depend on memory. That is why its title says 'relative validity', and the authors are clear that they measured agreement between methods, not accuracy.

Three more things are worth knowing. The studies disagreed with each other a great deal; for most nutrients, more than three quarters of the spread between results reflected real differences between studies and not chance, so the pooled figures are rough centres, not fixed values. The finding that questionnaires read out by an interviewer did a bit better rests on 6 of the 43 studies. And the summary at the top of the paper says the analysis shows 'digital' questionnaires are reliable, but the paper never says how many of the 43 were digital and none of its five subgroup comparisons sets digital against paper.

None of this is a reason to dismiss diet research. The researchers who run large studies know about these errors and have statistical ways to allow for them, and a pattern that turns up again and again in different populations does not hang on one form. It is a reason to read a single headline with some care. When a study reports that people in the top fifth for a food did better than those in the bottom fifth, the sorting into fifths was done with a tool like this one. It has the broad order about right, the exact amounts wrong, and for some nutrients not much to go on at all.