How Percentages Actually Work — And How They Get Misread.
Percentages compare a part to a whole, but the same true number can be framed to look alarming or trivial. Here's how to read one correctly.
- A percentage is always (part ÷ whole) × 100 — but two people can run that same formula on the same real event and produce numbers that feel completely different.
- 'Twice the risk' and 'an extra 0.01% chance' can both be true statements about the same statistic — relative risk and absolute risk answer different questions.
- You can't average two percentages unless the groups behind them are the same size — a 90% pass rate on 10 students and a 20% pass rate on 200 students is not a 55% combined rate.
- A margin of error of '±3 percentage points' means the true number could plausibly be anywhere in a 6-point band, which is wider than most headlines imply.
- A percentage calculated from a handful of cases — '100% of our test users loved it' — can be true and still tell you almost nothing.
01Foundations
The formula is the easy part
A percentage answers one question: how much of the whole is this part? Divide the part by the whole, multiply by 100, done. 25 correct answers out of 200 questions is 25 ÷ 200 × 100, or 12.5%. That arithmetic takes about five seconds, and running through discounts, reverse percentages, and the mechanics of the formula itself is exactly what the percentage calculator on this site is built for, with a full worked example and a dozen answered edge cases.
This article is about something the formula doesn’t warn you about: the same true percentage can be presented in ways that make it feel completely different, and most people never learn to spot the gap. A drug “doubling your risk” and a drug “adding a 0.05 percentage point risk” can be the exact same fact, stated two different ways.
A poll showing a candidate “up 3 points” might not mean anything at all. Reading a percentage correctly is a separate skill from calculating one, and it’s the more useful of the two once you’re out of a classroom and into a news feed, a product page, or a doctor’s office.
The same ambiguity shows up somewhere much smaller: a restaurant check. A 20% tip on a $60 bill is $12 — easy. But which $60? Receipts list a pre-tax subtotal and a post-tax total a couple of lines apart, and they’re not the same number. On a $60 subtotal with 8% sales tax, the total comes to $64.80. Twenty percent of $60 is $12; twenty percent of $64.80 is $12.96. Neither calculation is wrong — 20% of a number is always 20% of that number — but “tip 20%” quietly depends on which of those two numbers you started from, which is exactly the kind of question this article keeps coming back to.
02Relative vs. absolute
“Twice the risk” and “barely any risk” can both be true
The most consequential percentage trick isn’t a calculation error at all — it’s a choice about which of two legitimate percentages to lead with.
Absolute risk is the plain probability of something happening: 12 people out of 100, or 12%. Relative risk compares two probabilities to each other: how much bigger is one group’s risk than another’s. Both are real, correctly calculated percentages. They just answer different questions, and headlines almost always report the more dramatic one.
A commonly cited example: a medication was found to raise the annual risk of a seizure from about 0.057% to about 0.16%. Framed as relative risk, that’s roughly 2.8 times higher — “nearly triple the risk of seizures.” Framed as absolute risk, it’s an increase of about 0.1 percentage points, or roughly 1 additional case per 1,000 people per year. Neither number is wrong. One just sounds far more alarming than the other, for reasons that have nothing to do with how many people are actually affected.
"Nearly 3× the risk"
Technically accurate — 0.16% is roughly 2.8 times 0.057%. This is the number that makes headlines, because ratios of small numbers can look enormous even when neither number is meaningfully large on its own.
"About 1 extra case per 1,000 people"
Also accurate, and calculated from the exact same two numbers. This framing tells you what the risk actually means for an individual, which is usually the more useful question.
This isn’t a hypothetical concern. In October 1995, the UK Committee on Safety of Medicines warned that certain oral contraceptive pills were associated with roughly twice the risk of venous thromboembolism (blood clots) compared with older formulations. That relative-risk framing drove widespread media coverage and a wave of women stopping the pill or switching brands without medical advice.
A later meta-analysis put the relative risk increase in the same range — about 1.7 times — but translated it into an absolute increase of roughly 11 additional cases per 100,000 women per year: a real effect, but a small one in absolute terms. In the following year, use of oral contraception among girls under 16 in the UK fell from about 40% to about 27%, and researchers subsequently linked the scare to a measurable rise in unintended pregnancies.
The relative-risk number wasn’t false. It was just the half of the story that, on its own, led a lot of people to a decision the full picture might not have supported.
03Points vs. percent
A percentage point is not a percent
This distinction gets its own full treatment, with the formula and more examples, in the percentage calculator’s FAQ — but it’s worth flagging here because it’s the mechanism behind a specific style of misleading headline. If unemployment moves from 4% to 5%, that’s a rise of one percentage point, but a 25% increase relative to where it started (1 ÷ 4 × 100).
Both descriptions are technically correct. “Unemployment jumps 25%” describes the same real change as “unemployment up one point,” but reads as a far bigger story, which is exactly why the relative framing shows up more often in a headline than the absolute one does.
04Margins of error
Reading a poll: the margin of error is bigger than it looks
Every political poll comes with a margin of error, usually something like “±3 percentage points.” Most people read past it. It’s worth actually unpacking, because it changes what a poll is allowed to claim.
A margin of error of ±3 points at the standard 95% confidence level means: if the same survey were run 100 times on 100 different random samples, about 95 of those runs would land within 3 points of the true population value. It’s a band of plausibility around the reported number, not a guarantee of precision.
The part that trips people up is what happens when you compare two candidates, or two points in time, rather than reading a single number in isolation.
A single candidate at 48%, ±3 points
Their true support most likely falls somewhere between 45% and 51%. That is already a 6-point-wide range for one number.
Two candidates 3 points apart
If one candidate polls at 48% and the other at 45%, the gap between them carries roughly double the single-candidate margin — around ±6 points — because both numbers carry their own uncertainty. A 3-point lead is not reliably outside that range.
A shift across two separate polls
If a candidate's lead grows from 5 points to 8 points between two polls, that 3-point shift can fall entirely inside the combined margin of error of both polls, meaning the 'shift' may not be a real change at all.
±6 pts
Subgroup breakdowns inside a poll — a specific age group, region, or demographic — are worse, because they’re based on a smaller slice of the total sample. A subgroup that makes up 15% of a 1,000-person survey might only represent about 150 actual respondents, which can push its individual margin of error out to ±8 points or more. A headline built on “young voters swung 6 points” drawn from that kind of subgroup is very often reporting statistical noise as if it were a trend.
05Simpson’s paradox
When the total tells the opposite story of every group inside it
In the fall of 1973, UC Berkeley’s graduate division admitted about 44% of male applicants and only about 35% of female applicants — a gap large enough that the university was sued for sex discrimination.
Statisticians Peter Bickel, Eugene Hammel, and J. William O’Connell went back through the admissions data department by department to find where the bias was coming from, and found something unexpected: checked individually, most departments showed no bias against women, and in several of the largest ones, women were admitted at a higher rate than men.
The explanation was that women had applied in much greater numbers to the university’s more competitive departments — the ones with low admission rates for everyone, regardless of sex — while men applied more heavily to departments that admitted a larger share of all applicants. Pooling every department into a single overall number buried that pattern and made an aggregate gap look like direct bias, when the real driver was which departments each group chose to apply to.
The Berkeley case is the standard textbook example specifically because the reversal was so large and so well documented, but the same pattern shows up anywhere a single combined rate is used to compare groups that aren’t actually comparable in makeup — hospital survival rates across facilities that treat different severities of case, batting averages compared across a partial season, or approval ratings pooled across regions with very different underlying populations. The fix is the same every time: check the breakdown before trusting the total.
06Averaging percentages
You can’t average two percentages without knowing the group sizes
A related mistake shows up constantly in spreadsheets and reports: averaging two percentages as if they were interchangeable numbers, without accounting for how many cases each one represents.
Say one class of 10 students has a 90% pass rate, and a much larger class of 200 students has a 20% pass rate. Averaging the two percentages directly — (90 + 20) ÷ 2 — gives 55%, and that number is wrong for describing the combined group, because it silently treats both classes as equally important when one is twenty times the size of the other.
(90% + 20%) ÷ 2 = 55%
Treats both classes as if they contributed equally, even though the second class has twenty times as many students. This number does not describe any real group of people.
(9 + 40) ÷ (10 + 200) ≈ 23.3%
9 students passed out of 10, and 40 out of 200 — combine the actual counts first, then divide. This is the real combined pass rate across all 210 students.
The correct method is to go back to the underlying counts — 9 out of 10 passed in the small class, 40 out of 200 passed in the large one — add the counts together (9 + 40 = 49), add the totals together (10 + 200 = 210), and only then divide: 49 ÷ 210 ≈ 23.3%.
That’s a genuinely different number from the naive 55% average, and it’s the only one that actually describes the combined group of 210 students. Any time a report averages percentages from groups of different sizes without this step, treat the result with suspicion.
07Small samples
A percentage needs a sample size attached to mean anything
The last trap is the simplest and the easiest to miss: a percentage on its own says nothing about how many cases it’s built from, and a small enough sample can make almost any percentage possible.
“100% of surveyed users loved the redesign” sounds like unanimous approval. If the survey reached 4 people, that 100% figure would have swung to 75% the moment a single one of them changed their mind — a group that small can only ever land on multiples of 25%, and no amount of confidence in the number changes that. The same caution applies in the other direction: a “0% failure rate” on a new component means very little if it’s only been tested a dozen times.
| Sample size | Smallest possible move | What that looks like |
|---|---|---|
| 4 | 25 points | 100% → 75% |
| 10 | 10 points | 100% → 90% |
| 50 | 2 points | 100% → 98% |
| 1,000 | 0.1 points | 100% → 99.9% |
Neither extreme means small samples are useless — sometimes a small sample is all that’s available, and it’s still better than no data. The point is narrower: a percentage without its sample size is an incomplete sentence. Credible reporting states both, and a percentage that shows up alone, with no count behind it, is worth a second look before you repeat it.
08Marginal vs. effective rates
A 22% tax bracket doesn’t mean 22% of your income
One of the most common percentage misreadings has nothing to do with risk or polling — it shows up on a pay stub. Someone who lands in a 22% tax bracket often assumes the government takes 22% of everything they earned that year. Most income tax systems don’t work that way, because they’re marginal: each bracket’s rate applies only to the slice of income that falls inside it, not to the whole amount.
Take a simplified three-bracket system: 10% on the first $10,000 of income, 15% on the next $30,000, and 22% on anything above $40,000. Someone earning $50,000 doesn’t owe 22% of $50,000, which would be $11,000. They owe 10% of the first $10,000 ($1,000), plus 15% of the next $30,000 ($4,500), plus 22% of the remaining $10,000 ($2,200) — a total of $7,700. Divide that by the full $50,000 and the effective rate is 15.4%, well below the 22% bracket the earner is technically “in.”
15.4%
That gap between the marginal rate — what applies to the next dollar earned — and the effective rate — what’s actually paid across the whole income — is also why a raise that pushes someone into a higher bracket almost never results in less take-home pay overall. Only the income above the new threshold gets taxed at the higher rate; every dollar below it keeps being taxed exactly as before.
09Wrapping up
What to actually check
None of the examples above involve bad arithmetic. Every percentage in this article was calculated correctly by the formula everyone learns: part divided by whole, times 100. The gap between a true percentage and a misleading one almost never lives in the math — it lives in which percentage got chosen to represent the story, and what got left out.
Four checks cover most of it: ask whether “more likely” is relative or absolute, check whether a percentage point difference is being reported as if it were a percent difference, look for the sample size behind any percentage that sounds too clean, and be suspicious of a single combined percentage describing groups that might not be the same size underneath it.
None of that requires more than the same formula you already know — just a habit of asking what’s standing behind the number before you accept what it’s implying.
Sources: Bickel, Hammel & O’Connell, “Sex Bias in Graduate Admissions: Data from Berkeley,” Science 187 (1975); Pew Research Center, “Understanding the Margin of Error in Election Polls” (2016); NIH StatPearls, “Relative Risk”; UK Committee on Safety of Medicines 1995 advisory and subsequent meta-analysis on third-generation oral contraceptives and venous thromboembolism risk.
Frequently asked questions
What's the difference between relative risk and absolute risk?
Absolute risk is the plain probability of something happening — 12 out of 100 people, or 12%. Relative risk compares two probabilities to each other, like 'twice as likely.' A rare event can double in relative terms while barely moving in absolute terms: going from a 0.05% risk to a 0.1% risk is a 100% relative increase but only a 0.05 percentage point absolute one.
Why can't you just average two percentages together?
Because a percentage hides the size of the group it came from. Averaging 90% (from 10 people) and 20% (from 200 people) as (90+20)/2 = 55% ignores that the second group is twenty times bigger and should count for far more. The correct method is a weighted average: combine the actual counts first, then divide — which gives a very different, and correct, answer.
What does a poll's margin of error actually mean?
A margin of error of plus or minus 3 percentage points, at the standard 95% confidence level, means that if the same poll were run 100 times, about 95 of those times the result would land within 3 points of the true population value. Two candidates within a combined 6 points of each other are effectively tied, not clearly ahead or behind.
Why did Berkeley's 1973 admissions data reverse when broken down?
Berkeley admitted 44% of male applicants and 35% of female applicants overall in 1973, which looked like bias. But checked department by department, most individual departments favored women or showed no gap — women had applied in greater numbers to the more competitive departments, which pulled their overall rate down. This reversal pattern is called Simpson's paradox.
Why is '100% of users loved it' sometimes meaningless?
Because percentages don't carry their sample size along with them. If only 3 people tried something and all 3 liked it, that's a true 100% — and also weak evidence, since a single fourth person disliking it would drop the figure to 75%. The smaller the group behind a percentage, the more a single case swings the number, which is why credible statistics report the sample size next to the percentage.