Relative and absolute risk: how to read health headlines

Two ways of reporting the same result

Imagine two groups of people followed over time. Each group's absolute risk is the proportion of people who have the problem being studied in that period. The relative risk divides one group's risk by the other's; 1 minus that ratio is the relative risk reduction, the “30% lower” of the headlines. The absolute risk reduction is the difference between the two absolute risks, and the number needed to treat is the inverse of that difference: how many people would have to receive something for one of them to avoid the problem.¹

The five figures describe the same result, but only the absolute ones show where you start from. The clearest approach is to give the risks of both groups: from 20 to 14 in 1,000 over ten years.¹

The same 30% doesn't mean the same for everyone

A relative percentage doesn't carry the starting point within it. A relative risk of 0.5 is consistent with going from 20% to 10%, from 1% to 0.5%, or from 0.0004% to 0.0002%.³ All three are “half the risk,” but only in the first are 10 out of every 100 people spared the problem; in the last, the difference is tiny.

And the starting risk varies from one person to another. In studies from several countries, the classic cardiovascular risk factors were associated with similar relative risks in men from Northern Ireland, France, the United States and Germany, although the starting risk of coronary heart disease varied widely between those countries.⁵ With the same relative risk, the number of people affected isn't the same.

Why the relative figure is more impressive

A Cochrane review pooled 35 studies, with 83 comparisons, in which the same effect was presented in different ways to health professionals and to the general public. Compared with the absolute reduction, the relative one wasn't understood any better, but it made the effect seem bigger and was more persuasive. Compared with the number needed to treat, the relative reduction also seemed bigger and was more convincing. There were no differences between professionals and the public.²

In a telephone survey in New Zealand, three cancer screening tests with exactly the same benefit were described, with the benefit reported in three different ways. When the benefit was described as “getting tested every two years reduces the risk of dying from this cancer over the next ten years by a third,” 80% of 306 people said they would probably get tested; as “from about 3 to about 2 deaths in 1,000 over ten years,” 53%; and as “if about 1,000 people get tested every two years, one death is prevented every ten years,” 43%.¹

From the scientific paper to the headline

The bare relative figure isn't born in the newspaper. In a review of 222 articles with ratio measures in the abstract, published between 2003 and 2004 in six major medical journals, 68% didn't give in the abstract the absolute risks behind the first risk ratio they reported, and in about half of those articles the absolute risks weren't reported anywhere in the article. In cohort studies, the abstract gave the absolute risks in 21% of cases; in randomized trials, in 62%.³

Press releases don't always help either: according to an analysis cited in the review by Gigerenzer and colleagues, of the medical journal press releases that compared groups, 45% gave relative risks without the starting risk.¹ And what press releases say makes its way into the news. In 462 health press releases issued by 20 UK universities in 2011, 40% gave more direct advice than the paper; among observational studies, 33% spoke of cause more categorically than the paper itself; and among studies with animals, cells or simulations, 36% extended the results to people beyond what the paper said. When the press release exaggerated the causal relationship, 81% of the news stories did too; when it didn't, 18%.⁴

Association isn't effect: what's behind “40% lower”

Many figures linking a habit to living longer come from cohort studies: something is measured in a large group of people, who are then followed for years. Their result is usually a hazard ratio, which compares the rate at which the problem occurs in one group and in the other over the follow-up; it's a type of relative risk.⁶⁵

A real example: in a meta-analysis of 15 cohorts, the group taking about 5,800 steps a day had a 40% lower risk of death than the group taking the fewest steps. That 40% is 1 minus the hazard ratio, 0.60, adjusted for other factors.⁷ The main text of the study gives the starting point for the whole sample: 3,013 deaths among 47,471 people, followed for a median of 7.1 years; according to the authors, 10.1 deaths per 1,000 person-years of follow-up.⁷ The 40% doesn't mean “40 out of 100 people,” nor does it say on its own how many fewer deaths there were in that group. What those studies show, we explain in the article on how many steps a day.

In addition, in an observational study the difference between groups may be due to other factors that go along with the habit; this is called confounding.⁵ Researchers adjust for the factors they know about and measure, but the authors of the steps analysis themselves warn that, since the data are observational, causal inferences can't be made: residual confounding may remain, or reverse causation, that is, poorer health making people walk less rather than the other way around.⁷

Nor does it say how many more years people live. A methods article written for clinical trials explains that the hazard ratio doesn't translate directly into how long it takes for the problem to arrive;⁶ in our view, the same reasoning applies to mortality cohorts: “a 30% lower risk of death” isn't “living 30% longer.” And living more years isn't the same as living more years in good health: we explain the difference between those two indicators in life expectancy and healthy life years.

Five questions for reading a headline

You don't need formulas. These questions are our own inference from the sources of this article, not a validated checklist:

1. From how much to how much, and over what period? Look for both groups' risks, in people out of 100 or out of 1,000, and the time frame. If there's only a percentage, half the information is missing.¹²
2. Risk of what? Fewer deaths from a specific cause doesn't mean fewer deaths overall; fewer deaths from any cause is a better sign that something helps, and in a test of interpreting medical data only 20% of participants recognized this.¹
3. What kind of study? If it's observational, it shows an association that may be due to other factors.⁵ If it was done in animals or cells, its results don't simply carry over to people, even though press releases often carry them over.⁴
4. In whom? The age, sex, country and health of the people studied change the starting risk, and with it what the percentage means.¹⁵
5. And what isn't reported? If the benefits are given in relative terms and the harms in absolute terms, you're comparing figures in different currencies. In three major medical journals, between 2004 and 2006, one in three studies of treatments that reported benefits and harms measured them with different yardsticks, almost always in that direction.¹

These questions help with reading one headline. When what overwhelms you is a stream of messages that contradict each other, another article explains how to choose a few sources and when to stop searching.

A group figure doesn't tell you how a change works for you, either. If you want to check it in your own case, framing it as a question and a personal test helps more than looking for the percentage in a headline.