Is the 50th percentile the median?

Yes, with the chosen sample convention specified where needed.

Can a percentile value fall between observed values?

Yes, under interpolation methods.

Does the 99th percentile mean the maximum?

Generally no; the sample method and sample size determine the estimate.

AI Glossary / Statistics and probability / Intermediate

Percentile

Definition

A percentile is a value marking a specified cumulative position in a distribution. The 75th percentile, for example, corresponds to cumulative probability 0.75. In a sample, percentiles are estimated from ordered observations using a defined calculation rule. Different accepted rules can produce different values, particularly for small samples, and ties require care in interpreting the proportion below a value.

What it means

Percentiles describe position rather than an average. A high percentile of response time describes the slower part of the distribution; a high percentile of a performance score describes a relatively high score within the chosen reference population.
The direction of what is desirable depends on the variable. A percentile is also different from a percentage: a score of 80% is a measured proportion, whereas the 80th percentile is a location relative to a distribution.

How it works

Choose the population or dataset and identify the measurement whose distribution matters. Clean units and missing values, and specify whether observations receive equal weight. Survey weights or exposure weights require a corresponding weighted procedure.
Sort the valid observations from smallest to largest. Select the requested cumulative probability, such as p = 0.90 for the 90th percentile.
Apply a documented sample-quantile method. One method selects the observation at rank ceil(n times p), using one-based ranks for 0 < p <= 1, with the sample minimum used at p = 0. Other methods interpolate between neighboring ordered values using different formulas for the fractional rank.
For example, a common linear interpolation rule uses rank 1 + (n minus 1) times p. Another uses (n plus 1) times p with appropriate endpoint handling. Both refer to the same percentile concept but define different finite-sample estimates.
For a theoretical distribution, a standard quantile definition is the smallest value at which the cumulative distribution function reaches or exceeds p. In continuous strictly increasing distributions, that value has exactly probability p at or below it. With discrete values or ties, the cumulative probability can jump past p.
State the percentile method, reference period, sample size, and any weighting. A software setting is part of the result, especially when a threshold determines eligibility or service-level reporting.
Distinguish a sample summary from an estimate of an unknown population percentile. Tail estimates can be unstable when little data support them. Sampling uncertainty, dependence, and measurement limits should influence how confidently a threshold is used.
A percentile of pooled observations is generally not the average of subgroup percentiles. Reconstructing a combined percentile normally requires the underlying distribution or a suitable distribution summary, together with the intended weighting.

Three practical examples

AI operations: a team reports the 95th percentile of response latency to describe slower user experiences.
Workforce analysis: analysts locate a salary within a clearly specified comparison distribution.
Quality monitoring: engineers track upper-tail defect measurements instead of relying only on a mean.

Think of it this way

Think of lining up observations from smallest to largest and marking a position in that line. When the desired position falls between people, an interpolation rule decides how to describe the gap. The rule must be stated if the exact value matters.

Worked example

Suppose eight recorded response times, in milliseconds, are 10, 20, 30, 40, 50, 60, 70, and 80. We want the 75th percentile.
The nearest-rank rule uses ceil(8 × 0.75) = 6, returning the sixth value, 60 milliseconds. Six of the eight observations are at or below 60, while five are strictly below it.
The rule using 1 + (n minus 1)p gives rank 1 + 7 × 0.75 = 6.25. Interpolating one-quarter of the way from 60 to 70 returns 62.5 milliseconds.
The rule using (n plus 1)p gives rank 9 × 0.75 = 6.75. Interpolating three-quarters of the same gap returns 67.5 milliseconds.
These different answers are not necessarily arithmetic mistakes. They reflect different sample-percentile conventions applied to a deliberately small dataset. None is automatically the true population 75th percentile, and eight observations provide little information about more extreme tails.

Illustrative case study

Fictional case: Juniper Assist finds that two dashboards report different latency percentiles from the same eight test sessions. The analysts trace the discrepancy to interpolation settings. They standardize the reporting rule, preserve the original observations, and expand monitoring before using a tail percentile as a contractual service target.

Common misconceptions and mistakes

Do not say that exactly p percent of observations must be strictly below every reported pth percentile. Ties, discrete data, and interpolation can prevent that statement from being true.
Do not confuse percentile rank with percentile value. A rank is a relative position; a percentile value is in the measurement's units.
Avoid averaging daily 95th percentiles and labeling the result the monthly 95th percentile. Sample sizes and distribution shapes matter.
A percentile depends on its reference population. Changing that population can change someone's relative standing even when their measured value stays unchanged.

Business use

Percentiles help businesses describe service tails, compare distributions, and set operational thresholds. Reliable use requires a common definition across reports, enough data for the target tail, and attention to changes in customer or workload mix. A percentile should accompany context about the underlying distribution and the consequences of exceeding the threshold.

Check your understanding

Why do the three worked calculations differ?

Answer: They use different rank and interpolation rules.

Why is 60 milliseconds not a score of 75%?

Answer: It is a measured time located at a specified position in this sample's ordered distribution.

Related terms

Quantile
Median
Quartile
Order statistic
Cumulative distribution function
percentile rank
interpolation
tail latency
weighted quantile

Learn next

Learn next how empirical distribution functions handle ties, how quantile algorithms differ, and how uncertainty increases when estimating extreme population percentiles from limited observations.

Sources

Sources: NIST Dataplot, Percentile, https://itl.nist.gov/div898/software/dataplot/refman2/auxillar/percenti.htm ; Howell, Statistical Methods for Psychology glossary, https://www.uvm.edu/~statdhtx/methods8/Glossary/Glossary.html . R, Sample Quantiles, types 1, 6 and 7: https://stat.ethz.ch/R-manual/R-devel/library/stats/html/quantile.html . These primary references describe percentile position and alternative sample calculations. All response times and the fictional case are original examples.

Previous: Percentage of agreement Next: Percentile Bootstrap Confidence Interval