Is the median a fractile?

Yes. It corresponds to the 0.5 probability level.

Must a sample fractile appear in the data?

No. Interpolating methods can return a value between observations.

Do all software packages return identical quartiles?

Not necessarily. Their quantile conventions or defaults may differ, so the method should be specified.

AI Glossary / Statistics and probability / Foundation

Fractiles

Definition

Fractiles are values that locate specified fractions of a distribution, including quartiles, deciles and percentiles. The term is commonly used as a general label for quantiles. For a continuous distribution with a strictly increasing cumulative distribution function, the p-fractile is the value below which a fraction p of the probability lies. In discrete distributions or finite samples, ties and gaps require a convention, and the fraction strictly below a reported value need not equal p exactly. A fractile is a cutoff value, not the probability level itself.

What it means

Fractiles describe positions within an ordered distribution. Quartiles mark quarter-based positions, deciles tenth-based positions and percentiles hundredth-based positions. The median is the 50th percentile or 0.5 quantile. These summaries can reveal how a typical observation differs from one near the upper tail, without assuming a normal distribution. For example, median response time and the 95th percentile answer different operational questions. A high percentile is not the maximum, and the same percentile level can correspond to very different values in different populations.

How it works

For a population distribution, a common definition is Q(p) = inf{x: F(x) is at least p}, where F is the cumulative distribution function. For sample data, sort the observations and apply a specified quantile estimator. Some methods select an observed order statistic; others interpolate between neighboring observations. One common interpolation rule uses position h = 1 + (n - 1)p and interpolates between the surrounding ordered values. Different accepted methods can differ, especially for small samples or extreme percentiles. State the method, sample size, units and any weighting or filtering rules when reproducibility matters.

Three practical examples

Service monitoring: report the median and a high percentile of response times to distinguish typical experience from slower cases.
Distribution comparison: compare quartiles across groups while checking whether their sampling and measurement definitions match.
Data exploration: use the interquartile range to summarize the width between the first and third quartiles, alongside other information about the distribution.

Think of it this way

Think of arranging people from shortest to tallest and placing markers at selected positions in the line. A marker identifies a height associated with a fraction of the group. It is not that fraction of the tallest person's height, and tied heights can make the boundary less exact.

Worked example

Consider eight hypothetical observations [2, 4, 5, 7, 10, 12, 15, 21]. Using h = 1 + (n - 1)p, the first quartile has position 2.75, giving 4 + 0.75(5 - 4) = 4.75. The median has position 4.5 and equals 8.5. The third quartile has position 6.25 and equals 12 + 0.25(15 - 12) = 12.75. The interquartile range is 12.75 - 4.75 = eight. At p = 0.9, position 7.3 gives 15 + 0.3(21 - 15) = 16.8. Under the inverse empirical distribution convention, however, the 90th percentile selects position ceiling(8 times 0.9) = eight and equals 21. Neither result should be presented without its method when comparing software outputs. The example also shows that an interpolated sample fractile need not be an observed value.

Illustrative case study

Fictional case: Hawthorn Service Analytics receives two different 90th-percentile results for a small test dataset. Analysts trace the discrepancy to different quantile conventions rather than a calculation error. They agree on a documented method for the dashboard and retain the sample size beside the result. For broader reporting, they also examine uncertainty and changes in the mix of requests. This invented case illustrates why consistent definitions are essential before interpreting a percentile difference as a change in service quality.

Common misconceptions and mistakes

Do not confuse the 90th percentile with 90% of the maximum or with the mean of the largest 10%. Do not assume every reported percentile splits a finite sample into exact proportions, especially with ties. Extreme quantiles estimated from small samples are uncertain. A percentile rank and a percentile cutoff are related but different objects. Pooling data can change quantiles in ways that averaging group quantiles does not reproduce. Weighted surveys require an appropriate weighted method, and missing or censored observations may require more than sorting the available measurements.

Business use

Businesses can use fractiles to communicate variation that a single average conceals. Select percentile levels that correspond to meaningful questions and report units and measurement scope clearly. Preserve consistent calculation conventions across systems and time periods. When a high percentile informs a service target, assess whether the sample supports that tail estimate and distinguish observed performance from a guaranteed future level. Avoid ranking groups solely on small numerical differences without considering sampling variation and case mix. Fractiles are descriptive tools whose usefulness depends on the population and measurement process they summarize.

Check your understanding

Is the 75th percentile always three-quarters of the largest observation?

Answer: No. It is a position-based cutoff in the distribution, calculated using an appropriate quantile convention, rather than a fixed fraction of the maximum value.

Related terms

Learn next

Learn the distinction between distribution quantiles and their sample estimates. Then examine quantile conventions, weighted quantiles and uncertainty in tail estimates before using percentiles for comparisons or operational commitments.

Quantile

Sources

Howell, Statistical Methods for Psychology glossary, fractiles: https://www.uvm.edu/~statdhtx/methods8/Glossary/Glossary.html R statistical documentation, Sample Quantiles and nine algorithms: https://www.stat.ethz.ch/R-manual/R-devel/library/stats/html/quantile.html

Previous: Fractal surface Next: Fraction of successes