Math paradoxes

Glossary / Absolute and relative risk

Glossary · absolute and relative risk

A pill halves the risk. How many people does it help?

"Cuts the risk by 50%" sounds like rescue for half of us. But if two in a thousand fall ill without the pill and one with it, a thousand people swallow it and it helps one. Both descriptions are true. The first is relative risk: how many times the danger changed. The second is absolute: how many people it actually touches. Headlines love the first; decisions should rest on the second.

In plain wordspictures, analogies, small numbers

“In plain words” is hidden on all puzzles.

Pills that "double the risk"

October 1995, Britain. The Committee on Safety of Medicines sends an urgent letter to 190,000 doctors, pharmacists and public health directors and makes an emergency announcement to the media: third-generation contraceptive pills double, that is, raise by 100%, the risk of blood clots in the legs and lungs compared with the previous generation of pills.

Take 7,000 women on the new pills. How many of them get a clot precisely because their pills are new rather than old?

A thousand people and one headline

Each dot is a person. Choose how many in a thousand fall ill without the treatment and how the treatment changes the risk. Then switch the headline: the same result can be told as "minus 50%" or as "one in a thousand". Below is how many people must be treated to help one.

Relative change in risk–
Absolute change–
–

Rigoroustables, charts, formulas

“Rigorous” is hidden on all puzzles.

What absolute and relative risk are

Absolute risk

The share of people who meet the event over some period: fall ill, get a side effect, buy. It is said as "so many in a thousand" or as a percentage.

without: 2 in 1,000 = 0.2% · with: 1 in 1,000 = 0.1%
absolute difference: 0.2% − 0.1% = 0.1 percentage points

Relative risk

The ratio of two absolute risks: how many times bigger or smaller the risk is in one group than in another. The number of people cancels out of it; only "so many times" is left.

0.1% ÷ 0.2% = 0.5 → the risk is halved, "down 50%"

A third number joins the two and best answers "so what does it mean for me?": the number needed to treat (NNT), how many people must take the treatment for one of them to avoid the illness. It is one divided by the absolute difference: 1 ÷ 0.001 = 1,000. If the treatment raises the risk, the same number is called the number needed to harm (NNH).

The same "minus 50%", different worlds

Relative risk doesn't know where you started. The treatment is the same in every row: it halves the risk, a 50% cut. Yet its benefit differs by a factor of hundreds.

Ill withoutWithAbsoluteNNT
1 in 1,0000.5−0.05 pts2,000
10 in 1,0005−0.5 pts200
100 in 1,00050−5 pts20
300 in 1,000150−15 pts≈ 7

"pts" are percentage points: the difference of two percentages, not a percentage of a percentage.

Benefit in percent, harm in pieces

The slyest trick is to give the benefit as a relative number and the harm as an absolute one. "The drug cuts the risk of cancer by 48%, and clots occur in only 1% of those who take it" sounds like an obvious yes. Here is a real drug facts box for tamoxifen, from a trial of 13,000 women at high risk of breast cancer, over five years:

PlaceboDrugChangePer 1,000
Invasive breast cancer3317−48%16 fewer
Clot in a leg or lung510+100%5 more

Now both rows are in the same units. "Minus 48%" and "plus 100%" push in different directions, while 16 and 5 per thousand can honestly go on the same scales: for roughly every 63 women taking the drug, one fewer cancer; for every 200, one more clot. Which matters more is for the person to decide with a doctor. But you can only decide when both numbers are absolute.

Source: the "Prescription drug facts" box by L. Schwartz, S. Woloshin and H. G. Welch (2007), reproduced in G. Gigerenzer et al., Psychological Science in the Public Interest 8(2), 2008. Relative numbers and NNT are computed from it.

When to use which

The rule: relative risk tells how strong an effect is; absolute risk tells how much it matters to you.

  • Whether to take a drug or a vaccineAbsolute risk and NNT, for the benefit and the side effects separately. "By 50%" without the starting risk says nothing.
  • Whether an effect is strong in itselfRelative. It is often similar across groups with different starting risks, which is why studies compare treatments with it.
  • A headline saying "twice as dangerous"Ask "out of how many?". Doubling a one-in-a-million risk gives one in five hundred thousand.
  • An A/B test in marketing"Conversion +20%" is a relative change. From 5% to 6% it is just one point, that is, 10 extra purchases per thousand visitors.
  • Budgets and planningAbsolute. A hospital, an insurer or a warehouse needs to know how many cases there will be, not by what factor they grew.
  • Benefit against harmBoth in the same units, best in people per thousand. Mixed units are a sign someone is persuading you.

Three more strange consequences

  • Minus 50% and plus 1% are the same factA new operation lowers mortality from 2% to 1%. Deaths fell by 50%. But survivors went from 98% to 99%, only 1% more. A relative change depends on what you count: deaths or lives.
  • Relative risk has a ceilingIf an event already happens to 60% of people, it cannot become three times as common: there is no more than 100%. The largest possible rise is 1/0.6 ≈ 1.67 times. Loud "five times" claims only happen with rare events.
  • The rarer the event, the louder the percentageFrom one case in a thousand to two is +100%. The same "+1 in a thousand" on top of 300 cases is only +0.3%. Relative numbers inflate small risks and hide changes in big ones.
Deepergeneralisations, theory, history

“Deeper” is hidden on all puzzles.

The formulas

Let p₀ be the risk without the treatment and p₁ the risk with it (shares from 0 to 1).

relative risk RR = p₁ / p₀ · relative risk reduction RRR = 1 − RR

absolute risk reduction ARR = p₀ − p₁ = RRR · p₀ · NNT = 1 / ARR = 1 / (RRR · p₀)

The last formula is the whole paradox: the benefit to one person is the strength of the treatment times the starting risk. A headline reports only the first factor. For the current slider position above: RR = –, ARR = –, NNT = –.

The odds ratio is not the relative risk

Medical papers often show one more number, the odds ratio (OR). Odds are p / (1 − p), "how many cases per non-case". For rare events the OR almost equals the relative risk: from 2 to 1 in a thousand, OR = 0.4995 against RR = 0.5. For common events they part ways. If the risk falls from 40% to 20%, the risk has halved (RR = 0.5), but the odds went from 0.4/0.6 to 0.2/0.8, so OR = 0.375. Anyone who reads the OR as a relative risk sees "minus 62.5%" instead of "minus 50%".

A little history

The number needed to treat was proposed in 1988 by Andreas Laupacis, David Sackett and Robin Roberts in the New England Journal of Medicine, as a measure clearer to doctors than relative risk reduction. Gerd Gigerenzer, who made the 1995 British story a textbook case, points out it was not the first: similar pill scares happened in 1970 and 1977, and after each one the number of abortions rose. His advice to a news reader is short: whenever you hear "so many times", ask "how many in a thousand?".

Where it shows up

More on “Percentages and risk”

All terms →