Glossary / Percentage points
A website's conversion rate was 1% and is now 2%. The marketer is thrilled: "up a hundred percent!" The finance person shrugs: "up one." Both look at the same numbers, and both are right in their own way. The difference between two percentages has its own unit, the percentage point. Mix it up with a percent and you can blow news up a hundredfold or shrink it to nothing, and a headline will always pick whichever number sounds better.
“In plain words” is hidden on all puzzles.
Last month 1% of a website's visitors bought something; this month 2% did. The number of visitors is the same. Here are four headlines for the report. Mark every one you think is true and press "Check".
Only the most modest one is false: "up 1%" means from 1% to 1.01%.
A percent is always a share of something. "Up 1%" means up by a hundredth of what there was, that is, of a 1% conversion rate: plus 0.01. To say that from one percent to two is a step of one, you need a different unit, the percentage point. "Up 100%" and "up 1 percentage point" describe the same change in two languages: the first measures against the old value, the second simply subtracts one number from the other. In everyday speech people often say "up 1%" for "up 1 point", but then the reader has to guess what was meant, and the two readings differ a hundredfold.
Set the "before" and "after" values. At the top the same two numbers are read three ways; below them are a thousand people, one cell each: filled cells bought, fell ill or voted. And the generator underneath writes a loud, a calm and an honest headline for the same numbers.
Play with small values. From 0.1% to 0.2% is just one person in a thousand, yet the relative change is the same +100%. From 40% to 41% is already ten people, yet the relative change is only +2.5%. The relative change is the difference in points divided by the old value, so the smaller the base, the louder the same step sounds. That's why rare events (side effects, accidents, rare diseases) are almost always reported in percent rather than in points.
“Rigorous” is hidden on all puzzles.
A percentage point (pp for short) is the unit for the difference between two percentages. If a share was p₀% and is now p₁%, it changed by p₁ − p₀ percentage points. The percent in "up k%" is counted differently: against the old value. One formula links the two.
difference in pp = p₁ − p₀ · relative change = (p₁ − p₀) / p₀ · 100%
| Before → after | Difference | Relative change | Per 1000 people |
|---|---|---|---|
| 0.1% → 0.2% | +0.1 pp | +100% | +1 |
| 1% → 2% | +1 pp | +100% | +10 |
| 5% → 5.5% | +0.5 pp | +10% | +5 |
| 20% → 22% | +2 pp | +10% | +20 |
| 40% → 41% | +1 pp | +2.5% | +10 |
| 3% → 2% | −1 pp | −33.3% | −10 |
The rows 5% → 5.5% and 20% → 22% have the same relative change, +10%, but differences four times apart. The rows 1% → 2% and 40% → 41% have the same difference, +1 pp, yet relative changes 40 times apart.
Relative changes have another trap: each one is counted from its own "before". An item cost 100, in a sale it was cut by x%, and then raised by the same x%. It seems the price is back where it was. Move the slider.
The cut was counted from 100, but the rise from the already reduced price, so it is smaller in money. The order doesn't matter: a rise of x% followed by a cut of x% gives the same, 100 · (1 − x²), where x is the fraction. Minus 50% and plus 50% leave 75, and to get back after minus 50% you need plus 100%. The same goes for share prices, salaries and customer counts: a 50% fall is "won back" only by doubling.
“Deeper” is hidden on all puzzles.
Relative changes don't add up: −50% and +50% is not zero but −25%. And they aren't symmetric: a 50% fall cancels a 100% rise. Economists sometimes measure change in log points: 100 · ln(after / before). For small changes it's almost the same as percent (+1% ≈ +0.995 log points), and for large ones it behaves more honestly.
−50%: 100 · ln 0.5 ≈ −69.3 · +50%: 100 · ln 1.5 ≈ +40.5 · together −28.8 = 100 · ln 0.75
Log points add up: the sum of two steps equals the step from start to finish. And they are symmetric: doubling is +69.3 and halving is exactly −69.3, so you can see that one cancels the other. For shares close to 0% or 100% statisticians go further and compare odds (p / (1 − p)), because a share can't go beyond 0 to 100%, while odds can.
The word "percent" means two different things: a share of a whole ("2% of visitors bought") and a change relative to the old value ("sales rose 2%"). When the quantity itself is a share, the meanings collide, and "conversion rose 2%" becomes ambiguous. The percentage point keeps them apart: points for a difference between shares, percent for a change relative to itself. That's why A/B test reports write "+0.5 pp (lift +10%)", and statistics offices report unemployment changes in points.
The share of people who meet an event: for example, 2 in 1,000, or 0.2%.
A pill that halves the risk may help only one in a thousand of the people who take it.
How many times bigger or smaller the risk is in one group than in another: 1 in 1,000 against 2 in 1,000 is half.
Mortality falls from 2% to 1%: 50% fewer deaths, yet only 1% more survivors.