The Empirical Rule (68-95-99.7) Explained With Examples
The empirical rule, in one breath: if your data is bell-shaped, about 68% of values fall within 1 standard deviation of the mean, about 95% within 2 standard deviations, and about 99.7% within 3.
That's it. Three numbers. But those three numbers let you estimate probabilities in your head faster than most people can open a calculator app — which is why teachers drill them and why researchers sanity-check results against them.
The rule in z-score language
Since z-scores count standard deviations, the rule becomes:
- z between −1 and +1: 68.27% of values
- z between −2 and +2: 95.45% of values
- z between −3 and +3: 99.73% of values
The decimals matter if you need precision: it's 68.27, not exactly 68, and 95.45 rather than a clean 95. (That gap is why the exact 95% cutoff needs z = 1.96 instead of 2.) Outside ±3 SD, only about 0.27% of values remain — roughly 1 in 370. If your data claims lots of values out there, the data probably isn't bell-shaped, or something's wrong with the measurements.
Because the curve is symmetric, you can halve everything. Within ±1 SD lives 68.27%, so each shoulder — say, between z = 0 and z = +1 — holds about 34.13%. And the band between 1 and 2 SD holds (95.45 − 68.27) ÷ 2 = 13.59% per side. Stack a few of these and you can estimate almost any middle-of-the-curve probability without a table. For exact values, use the z score to percentile calculator.
Worked example: adult heights
Suppose heights in a large group of adults are roughly normal with mean 170 cm and standard deviation 10 cm. (Simplified numbers, chosen so the arithmetic stays clean — real groups differ by sex and population.)
- 68% of people measure between 160 cm and 180 cm (170 ± 10).
- 95% measure between 150 cm and 190 cm (170 ± 20).
- 99.7% measure between 140 cm and 200 cm.
Now a question teachers love: roughly what share of people are taller than 190 cm? Height 190 is z = (190 − 170) / 10 = +2. Outside ±2 SD there's 100 − 95.45 = 4.55%, half of it in the top tail: about 2.28%. If you meet 100 random adults from this group, expect only two or three to top 190 cm. Sanity check with the z score calculator: z = 2.00 → 97.72nd percentile → 2.28% above. The head-math and the exact answer agree.
Worked example: exam scores
A large exam has mean 500 and SD 100 (the old SAT section scale, roughly). The rule instantly tells you:
- About 68% of students scored 400–600.
- About 95% scored 300–700.
- A score of 750 sits at z = +2.5 — beyond the 95% band, inside the 99.7% band, and clearly rare: the exact figure is the 99.38th percentile.
Test-prep tutors use this constantly to translate a student's score into "how many students are ahead of me" without any tables.
The one condition everyone forgets
The empirical rule is a property of the normal distribution — the standard normal distribution is its purest form. If your data is skewed, the rule breaks, sometimes badly.
Income is the standard cautionary tale. Income has a long right tail of very high earners, so "within 1 SD of the mean" doesn't contain anything like 68% of households, and values more than 1 SD below the mean can be mathematically impossible (income can't go below zero). Applying 68-95-99.7 there produces confident nonsense.
Quick field test before you use the rule: Is a histogram roughly mound-shaped? Are the mean and median close? If yes, proceed. If the data leans to one side like a ski slope, stop.
Where the rule shows up in real life
Quality control is the big one. Manufacturers treat ±3 SD as the "voice of the process": parts landing outside it should occur only 0.27% of the time naturally, so an outside part signals a machine problem rather than bad luck. Six Sigma programs build an entire philosophy on pushing variation inside ever-tighter bands.
It's also the fastest way to interpret z-scores in conversation. Someone says "the result was 2.2 standard deviations out" and you immediately know: unusual, outside the 95% band, worth a second look. Practice on real problems until that reflex is automatic.
Frequently asked questions
What percentage of data falls within 2 standard deviations?
About 95% — more precisely 95.45% under a perfect normal distribution. That leaves roughly 4.55% (often rounded to 5%) outside, split between the two tails at about 2.28% each. This near-5% figure is exactly why ±1.96 (the exact 95% point) became the standard significance cutoff.
Does the empirical rule work for all data?
No — only for data that's approximately normally distributed (bell-shaped and symmetric). It fails badly for skewed data like incomes or lopsided counts. A rough check: if the mean and median are close and a histogram looks like a mound, the rule is a fair guide. Otherwise, don't apply it.
What percentage lies between 1 and 2 standard deviations?
About 13.59% on each side of the mean. The middle ±2 band holds 95.45% and the middle ±1 band holds 68.27%, so the ring between them holds (95.45 − 68.27) ÷ 2 = 13.59% per side. It's a handy number for estimating probabilities without a table.
Why is it called the empirical rule?
Empirical means observed from data rather than proven from pure theory. Long before exact tables existed, statisticians kept noticing that real bell-shaped datasets split roughly 68/95/99.7 around the mean. The percentages are properties of the normal curve itself, so the rule is really a memorable summary of that curve.