A headline in front of me says a thing doubles your risk. The paper it is describing reports that the outcome occurred in about one person in ten thousand in the unexposed group and about two in ten thousand in the exposed one.
Both statements are true. They are also, for the purpose of a person deciding what to do on a Tuesday, almost unrelated to each other.
Start from what a relative risk is. It is a ratio: the rate in one group divided by the rate in another. A ratio is scale-free by construction — that is the whole point of it — which means it deliberately throws away the one piece of information a reader needs, which is how big the thing was to begin with. Two in ten thousand and two in ten are both a doubling. The word carries no information about which one you are in.
The absolute risk is the other number: how many people per thousand, before and after. It is in the paper. It is nearly always in the paper. It is very often not in the sentence anybody reads.
The reason it goes missing is not that anyone is hiding it, and I want to be exact about this because the version of this complaint that assumes bad faith is both common and wrong. Relative risk is the correct measure for a great many scientific purposes. If you are asking whether a mechanism is real, the ratio is what you want, because it is comparable across populations with different baselines — a doubling in Finland and a doubling in Kerala are the same finding about biology, while two-per-thousand and eight-per-thousand are not. Researchers report ratios because ratios are what the science is about.
The trouble is that a ratio answers a question about the world and a person is asking a question about themselves, and nothing in the pipeline from paper to press release to headline is responsible for the translation.
Here is a way to hold onto it. A relative risk tells you how much the needle moved. An absolute risk tells you how big the dial is. You cannot act on the first without the second, and every sentence containing only the first is a sentence that has been rendered unusable while remaining entirely accurate.
There is a symmetric failure that gets less attention and is worth as much. A small relative risk on a very common outcome can be enormous — a four percent increase in something that happens to one person in three is a public health event, and it will be reported, if at all, as a modest association. The two errors point in opposite directions and have the same cause.
What I would ask for is one sentence in the press release, and it is not a hard sentence to write: about X in 10,000 people, rising to about Y. The paper already contains both numbers. Adding them costs nothing and removes the reader's only real defence, which at the moment is to assume every reported doubling is trivial — a heuristic that is right most of the time and is going to be badly wrong exactly once.
