How to read a paper

Every paper starts with an abstract: usually what the question was, what they did, what they found, and what they think it means, in that order. It is free to read even when the paper is not. The summaries on this map are written from abstracts, so if you read nothing else, read that.

The last two sentences of an abstract are the authors' own conclusion. Read those first, then go back and check whether the results sentence before them actually supports it. Sometimes it doesn't.

Look for the number of people, usually written as n = something. Twenty people is a pilot. Two hundred is a proper study. Twenty thousand is a registry or a national database. A finding in 20 people is a lead; a finding in 20,000 is a pattern.

Then check who they were. A study of Korean men with advanced disease may not describe a young woman recently diagnosed in Brisbane. That doesn't make it wrong; it makes it specific.

If the abstract mentions mice, rats or cells in a dish, the paper is about biology, not treatment. Many things that work in mice never work in people, and the gap is usually years even when they do. These papers are on the map because they explain mechanisms, which is useful, but no mouse result should change what you do.

A randomised controlled trial splits people by chance into a treatment group and a comparison group. It is the only design that can show a treatment causes an outcome. Everything else, cohorts, registries, cross-sectional surveys, shows that two things go together, which is weaker. Diet studies on this map are almost all of the second kind; drug trials are the first.

Within trials, the comparison matters. Against placebo tells you the drug does something. Head-to-head against another drug tells you which is better, which is the question you usually have.

A p-value below 0.05 means the result would be unlikely to appear by chance if the treatment did nothing. It does not mean the effect is large, or important, or will happen to you. A tiny effect in a huge study can have a very small p-value and still not matter.

Look instead for the size of the effect in plain units: how many more people reached remission, how many points a score moved, how many fewer flares. And look for the confidence interval, the range the true answer probably lies in. If that range includes zero, the study didn't really find a difference.

A 40 percent higher risk sounds alarming. If the baseline risk is 2 in 100, a 40 percent increase takes it to about 3 in 100. Both numbers are true; only the second tells you how much it matters to one person. When a paper gives only the relative figure, be suspicious of the headline.

A preprint has not been checked by other scientists yet. The map marks them. Treat them as promising, not settled.

A review summarises other papers; a systematic review does so with stated rules, and a meta-analysis pools their numbers. A good meta-analysis of several trials is the strongest evidence that exists. A narrative review in a minor journal is one person's opinion with references.

Near the end of most papers is a funding statement and a conflicts-of-interest list. Drug-company funding does not make a trial wrong, but it does mean the question was chosen by the company. Independent registries and government-funded cohorts have no product to sell, which is one reason the map leans on them for long-term safety questions.