The short answer: the Pth percentile is the value at or below which about P% of the data fall. For sorted data, a common position is i = (P/100)(n − 1) + 1, with linear interpolation between neighbours. For 12, 15, 18, 22, 25, 28, 31, 35, the 25th percentile is 17.25 and the 90th is 32.2.

Key takeaways

  • A percentile is a position, not a percentage score. The 90th percentile is not “90% correct.”
  • There is no single universal formula. Nearest-rank, inclusive, and exclusive methods give different values on small samples.
  • Interpolation assumes the gaps between observations are meaningful. For discrete data, a neighbouring observed value may be more honest.
  • Extreme percentiles need large samples. A 99th percentile from 50 observations is barely estimable.
  • Percentiles are not additive or averageable. Averaging two percentile values rarely gives a meaningful percentile.
  • Always state the reference population and the method used.

What a percentile means

The Pth percentile of a distribution is a value such that approximately P% of observations are less than or equal to it. The median is the 50th percentile, the first quartile is the 25th, and the third quartile is the 75th. A child at the 80th percentile for height is taller than about 80% of children of the same age and sex in the reference population.

The comparison is always relative to a specific group. Being at the 80th percentile nationally is different from being at the 80th percentile in a selective school. A percentile without a reference population is an empty statement, which is why growth charts and test norms specify the exact comparison group and date.

The rank formula and interpolation

Sort the n observations from smallest to largest. A widely used position formula is:

i = (P ÷ 100) × (n − 1) + 1
Percentile = x_floor(i) + (i − floor(i)) × (x_ceil(i) − x_floor(i))

The result i is a 1-based position, not an index into a zero-based array. If i is a whole number, the percentile is the observation at that position. If i falls between two positions, interpolate linearly: take the lower value plus the fractional part times the gap to the upper value.

This is the inclusive method used by many spreadsheets and statistical packages. Other definitions exist, and the choice matters most for small samples and extreme percentiles.

Worked example with eight values

Take the sorted data 12, 15, 18, 22, 25, 28, 31, 35, so n = 8. For the 25th percentile, i = 0.25 × 7 + 1 = 2.75. The value at position 2 is 15, the value at position 3 is 18, and the fraction is 0.75, so P25 = 15 + 0.75 × 3 = 17.25.

For the median, i = 0.50 × 7 + 1 = 4.5, so P50 = 22 + 0.5 × (25 − 22) = 23.5, matching the average of the two middle values. For the 90th percentile, i = 0.90 × 7 + 1 = 7.3, so P90 = 31 + 0.3 × (35 − 31) = 32.2.

The 25th percentile is below the third-smallest observation, and the 90th is above the second-largest. That is expected: percentiles describe position, and with only eight values each observation carries a lot of weight.

Why methods disagree

The nearest-rank method computes the smallest observation whose rank is at least P% of n. For P25 with n = 8, that is position ⌈0.25 × 8⌉ = 2, giving 15 rather than 17.25. The exclusive method uses i = (n + 1) × P/100, giving 2.25 and a value of 15.75. Three defensible methods, three different answers.

The differences shrink as n grows and vanish in the interior of a large, dense sample. They matter for small samples, for discrete data, and at the tails. Report the method, especially when a percentile is used as a cutoff for eligibility or payment. Software defaults are not a justification; they are a choice that should be documented.

For ordinal data such as ratings from 1 to 5, interpolation can produce a value like 3.7 that no respondent could give. In that case, report the observed category or use the nearest-rank definition.

Percentile versus percentage and percentile rank

A percentage is a fraction of a whole, such as 45 correct out of 50, or 90%. A percentile is a position in a distribution. Scoring 90% on a test says nothing about your percentile; if the test was easy, 90% might be the 40th percentile, and if it was hard, 70% might be the 95th.

Percentile rank runs the other way: it reports the percentage of observations at or below a given value. The value 25 in the example has three observations below it, so its rank is 3/8 = 37.5% under a simple “percent below” definition, or 43.75% under a midpoint definition that counts half of the ties. State which definition you mean, because “percentile” and “percentile rank” are routinely confused.

Percentiles, quartiles, and the IQR

Quartiles are specific percentiles: Q1 = P25, Q2 = P50, and Q3 = P75. The interquartile range is IQR = Q3 − Q1, the width of the middle half of the data. For the eight-value example, Q1 = 17.25 and Q3 = 0.75 × 7 + 1 = 6.25, so Q3 = 28 + 0.25 × 3 = 28.75 and the IQR is 11.5.

The IQR is resistant to outliers because it ignores the extreme quarters. That makes it a better summary than the range for skewed data, though it discards information and is less efficient than the standard deviation when the data are normal. Percentiles at 10 and 90 are also used to describe spread while excluding the tails.

Percentiles from a normal distribution

When a variable is approximately normal with mean μ and standard deviation σ, a percentile can be read from the standard normal curve. Standardize with z = (x − μ) ÷ σ and take the cumulative area Φ(z). With μ = 100 and σ = 15, a value of 120 has z = 1.333 and Φ(1.333) ≈ 0.9088, the 90.9th percentile.

The same conversion works in reverse: the 95th percentile is μ + 1.645σ = 124.7, and the 5th is μ − 1.645σ = 75.3. The 50th percentile is the mean, and the 16th and 84th are approximately μ − σ and μ + σ. This is why z-scores and percentiles are often treated as interchangeable — but only when the normal model fits.

Sample size and the tails

Percentiles near the extremes are hard to estimate. Just to expect one observation beyond the 99th percentile, you need about 100 observations; to estimate it with any stability, you need several hundred. A P99 from n = 50 is essentially the maximum of the sample and will jump around from sample to sample.

A rough planning rule is n ≥ 100 ÷ (100 − P) merely to expect one observation in the tail. For P95 that is 20, for P99 it is 100, and for P99.9 it is 1,000. Confidence intervals for extreme percentiles are wide even when the point estimate looks precise, so report the sample size and an interval when the stakes are high.

Common mistakes

  • Reading a percentile as a percent score. The two measure completely different things.
  • Averaging percentiles across groups. The average of two group percentiles is not the combined percentile.
  • Comparing percentiles from different tests or norms. A 90th percentile on one norm is not a 90th on another.
  • Using unsorted data. The rank formula assumes the observations are ordered.
  • Interpolating ordinal categories. A percentile of 3.7 on a 1–5 scale may be meaningless.
  • Ignoring the method. Nearest-rank and inclusive definitions can differ by a full observation on small samples.

Practical caveats

Percentiles compress a distribution into one number and hide its shape. Two groups can share the same median and differ completely in their tails. A change in the 90th percentile may reflect a shift in the whole distribution or only in the top decile, and the percentile alone cannot tell you which.

Measurement error also matters. A test with a standard error of 3 points cannot reliably distinguish the 80th from the 84th percentile. Growth charts address this with smoothed reference curves and z-scores rather than raw ranks. When a percentile drives a decision, ask how stable it is across plausible samples and whether a confidence interval crosses the cutoff.

Frequently asked questions

Is the 50th percentile the same as the median?

Yes. The median is the 50th percentile by definition, although different interpolation methods can give slightly different values for even-sized samples.

Can a percentile be 100?

With the inclusive formula, the 100th percentile is the maximum value. Some definitions treat the 100th percentile as undefined because no observation can exceed it. Either way, P100 is a boundary case.

Why do different calculators give different percentiles?

They use different rank definitions: nearest-rank, inclusive, or exclusive. The underlying data are the same; the interpolation rule differs.

What percentile is one standard deviation above the mean?

For a normal distribution, μ + σ is about the 84th percentile and μ − σ is about the 16th. This follows from Φ(1) ≈ 0.8413.

How many observations do I need for a reliable 99th percentile?

At least 100 just to expect one observation beyond it, and several hundred for a stable estimate. Report the sample size alongside the percentile.

Can percentiles be averaged?

No, not meaningfully. Percentiles are positions, not values with additive properties. To combine groups, pool the raw data and recompute the percentile.

Further reading

Keep exploring

Read next Logarithm: Base, Product Rule, and Scale A logarithm is an exponent in disguise. Learn the definition, the product, quotient and power rules, change of… Read next Order of Operations: PEMDAS and Ambiguous Expressions PEMDAS is a shared convention, not a law of mathematics. Learn the real precedence rules, the exponent and… Read next Significant Figures: Counting, Rounding, and Precision Significant figures record how precisely a value was measured. Learn the counting rules, the addition and multiplication rules,…