Glossary

Participants are split between groups by chance, one group receives the intervention, and the results are compared. Random allocation spreads things like age, diet and health status evenly across the groups, so a difference at the end can reasonably be attributed to the intervention.

Being randomised does not by itself make a result trustworthy. With twenty participants or two weeks of follow-up, randomisation still leaves very little you can conclude.

Results from several studies asking the same question are combined statistically into a single estimate. A direction invisible in one small study can appear once several are pooled.

Pooling does not improve the ingredients. Combine many poor studies and you get a poor conclusion. That is why this site reports how many studies were pooled, how many participants in total, and how the authors rated the certainty of the evidence.

A search of the literature on a question carried out to rules fixed in advance: which databases, which inclusion and exclusion criteria, which appraisal method, all declared before the work begins and then followed.

That is what separates it from citing whatever studies you like. Fixing the rules first makes it hard to select your way to the conclusion you wanted.

Keeping people unaware of who received the real intervention. If only the participants are unaware it is single-blind; if participants and researchers both are, double-blind.

The reason is that expectation moves results, especially for outcomes measured by questionnaire such as pain or sleep quality. So when a trial's main outcomes are questionnaires, this site notes if it was only single-blind.

The same people go through both conditions, in different orders. Each person serves as their own control, so differences between people drop out and smaller numbers can still detect an effect.

In exchange, a washout period sits between the conditions so the first does not carry into the second. Whether that gap was long enough is the thing to look at in this design.

The diet or habits of many people are recorded and the group is followed for years or decades to see who develops what. Nobody is randomly assigned, so this is observation rather than intervention.

That means it can show association but struggles to show cause. People who eat more nuts also tend to exercise more and smoke less. Statistical adjustment helps, but what was not measured cannot be adjusted for.

In controlled feeding the researchers design and supply the whole diet; in free living people eat as usual with one thing added or swapped.

The distinction often decides the conclusion. Controlled feeding fixes not only what went in but what came out. That is why a diet where nuts displaced saturated fat and a trial where nuts were simply added on top can give opposite answers.

It means the difference observed is hard to explain by chance alone. It does not mean the effect is large, and it does not mean it matters.

With enough participants, a tiny difference becomes significant. With too few, a real difference fails to reach significance. So read the size of the difference alongside the word, never the word alone.

The probability of seeing a difference this large in this data if there were really no difference at all. By convention, below 0.05 is called significant.

That 0.05 is a convention, not a line in nature. There is no meaningful gap between p=0.049 and p=0.055. When this site reports p=0.055 as a tendency rather than a finding, it is keeping to that convention rather than bending it.

The range in which the true value plausibly lies. It tells you more than the single estimate does.

Two things to look at. If the range crosses zero, no effect has not been ruled out. If the range is wide, the data failed to pin the estimate down. A diastolic result running from −13.61 to −2.02 is not something a single number can summarise.

How far apart the two group averages ended up, in the original units: mmHg for blood pressure, mmol/L or mg/dL for cholesterol.

Because the units survive, the size is easy to judge. That is why this site also converts figures into mg/dL. An LDL drop of 0.47 mmol/L lands differently once you read it as about 18 mg/dL.

One group's risk divided by the other's. A value of 1 means no difference; 0.81 means 19% lower.

That 19% is relative. Without knowing what the underlying risk was, the number alone cannot tell you how much this matters. Read it together with the entry on relative and absolute risk below.

This is where health writing misleads people most often.

If risk falls from 2 in 1,000 to 1 in 1,000, that is a 50% reduction in relative terms and 1 person per 1,000 in absolute terms. The same fact, two ways of writing it, and the bigger-sounding one usually goes in the headline. When you see only a percentage, look first for how many out of how many.

How much the studies pooled in a meta-analysis disagree with one another. Reported as I², and the higher it is, the more the studies are telling different stories.

High heterogeneity thins out the meaning of the pooled average. Some studies fell sharply, others rose, and reporting only the average amounts to saying it fell.

An analysis reporting that although the whole group showed no difference, some subset did. Planned before the study starts it is a subgroup analysis; carried out after seeing the results it is post-hoc.

Slice data enough ways and somewhere a difference appears by chance. So a post-hoc finding is treated as a hypothesis for the next study, not a conclusion. If the overall result was not significant, that fact comes first.

A four-step rating — high, moderate, low, very low — of how much confidence the researchers place in their own estimate. It judges the trustworthiness of the estimate, not the size of the effect.

Low means the true value may be substantially different from the estimate. So when a figure is large but the certainty is low, this site says so. The alternative is to erase the authors' own rating and keep only their number.

The chance that a study's design or conduct pushed the result in one direction. It is assessed item by item: whether allocation was properly randomised, whether blinding held, how dropouts were handled.

High risk of bias does not mean the result is wrong. It means the result cannot simply be taken at face value.

A finding that the treated and untreated groups did not differ. These are often treated as failed studies, so they are published less and cited less.

But no effect is also an answer. That is why this site reports trials where nothing moved, such as the 2023 macadamia trial or the 2020 cod supplement trial. Collect only the results that moved and the food looks more capable than it is.

The unit used for vitamin A. Vitamin A arrives in several forms — retinol from animal foods, beta-carotene from plants — and the body uses them with different efficiency, so RAE converts them onto one scale.

That is why the beta-carotene figure for a carrot and the retinol figure for liver cannot simply be added. They only compare once both are in RAE.

A score for how fast a food raises blood glucose, measured against glucose itself.

Two cautions. First, GI is defined on 50 g of carbohydrate from that food, which is rarely the portion anyone eats. Second, a low GI is not permission to eat more. The same food varies widely with cultivar, ripeness and cooking.

The category of sugar the WHO advises limiting: sugars added to food, plus those in honey, syrups and fruit juice.

There is an important exception. Sugars held inside whole or dried fruit are not counted as free sugars. That is not permission to eat without limit — it means they cannot be counted identically to table sugar.

A collective name for short-chain carbohydrates that are poorly absorbed in the small intestine and ferment further down: fructose, lactose, oligosaccharides and polyols.

For most people they cause no trouble. But with irritable bowel syndrome they readily produce gas, cramping and diarrhoea, so this site notes it on the ingredients concerned. The same food can move into the high category as the portion grows.

A fatty acid with no double bonds is saturated; with one, monounsaturated; with two or more, polyunsaturated. That structural difference decides whether it sets at room temperature and how fast it goes rancid.

When reading, what matters is less the total than what replaces what. Trials that moved lipid numbers usually replaced saturated fat with unsaturated; trials that only added fat on top of the usual diet commonly moved nothing.

Both measure insulin resistance. HOMA-IR is an indirect index calculated from fasting glucose and fasting insulin, which makes it convenient. The clamp infuses insulin at a fixed rate while holding glucose normal and measures directly how much glucose that takes.

The clamp is the reference method in this field. It is laborious, so numbers tend to be small — but twenty people measured by clamp and twenty measured by fasting values do not carry the same weight.