Must groups have equal sizes?

No; the calculation can accommodate different sizes.

Is it a test of medians?

No. Medians may be used to center observations, but the comparison concerns spread.

Does a large p-value prove equal variability?

No. It indicates insufficient evidence against the tested null under the chosen calibration and assumptions.

AI Glossary / Statistics and probability / Advanced

Fligner-Killeen test

Definition

The Fligner-Killeen test is a rank-based procedure used to investigate whether independent groups have a common population spread, commonly described as a test of equal variances. A widely used version centers observations on their group medians, ranks the resulting absolute deviations and converts the ranks into normal scores. It is designed to be less sensitive to departures from normality than the classical normal-theory variance test. It is not a general test of equal means, and robustness does not mean that arbitrary dependence or distributional differences can be ignored.

What it means

Two processes may produce similar typical outcomes yet differ substantially in consistency. A spread comparison asks whether the observed dispersion differences are compatible with a common-scale explanation under the test's assumptions. Rank-based scoring limits reliance on the numerical size of extreme deviations. However, interpreting rejection specifically as a variance difference is clearest when the groups have comparable distributional shapes apart from location and scale. If shapes differ substantially, a result may reflect broader differences in their centered absolute-deviation distributions. Failure to reject is not proof of equal variances.

How it works

For the median-centered version, subtract each group's median and take absolute values. Pool these deviations across groups, assigning average ranks to ties. For total sample size N and rank r, form the score Phi-inverse[0.5 + r/(2(N + 1))], where Phi is the standard normal cumulative distribution function. Compare each group's mean score with the pooled mean, weighting squared differences by group size. Divide their sum by the pooled sample variance of the scores. The conventional large-sample reference is chi-square with k minus one degrees of freedom for k groups. Specify the centering convention, missing-data handling and calibration used; different choices need not give identical results.

Three practical examples

Process monitoring: compare consistency of measurements from independently sampled production settings.
Service evaluation: investigate whether response-time spread differs across groups while examining skewness and sampling structure.
Statistical diagnostics: assess evidence about unequal spread as part of a broader analysis, without making a preliminary test the sole basis for choosing all subsequent methods.

Think of it this way

Think of comparing how far shots land from each team's own center. Rank the distances across teams rather than comparing the centers themselves. A team that repeatedly receives high distance ranks may be less consistent, even if its typical location is similar.

Worked example

Take three hypothetical groups: [-2, -1, 0, 1, 2], [-4, -2, 0, 2, 4] and [-8, -4, 0, 4, 8]. All medians are zero. Among the 15 pooled absolute deviations, values zero, one, two, four and eight receive average ranks two, 4.5, 7.5, 11.5 and 14.5. Applying the normal-score formula gives group mean scores approximately 0.425954, 0.712908 and 1.132844. The pooled sample variance of all scores is approximately 0.254447. The weighted between-group score calculation gives a statistic of 4.967533. Using the conventional chi-square reference with two degrees of freedom, the upper-tail probability is exp(-4.967533/2), approximately 0.08343. At a prespecified 5% threshold this does not reject the null. These deliberately small groups illustrate the calculation; the asymptotic p-value is not an exact small-sample probability, and nonrejection does not establish equivalence.

Illustrative case study

Fictional case: Cedar Quality Analytics compares measurements from three settings after checking that observations represent independent experimental units. Its analyst plots the distributions, reports the median-centered Fligner-Killeen result and discusses the practical size of the observed spread differences. When a manager asks whether a nonsignificant result proves identical consistency, the analyst explains the limits of the sample and the distinction between a null-hypothesis test and an equivalence claim. The invented case illustrates interpretation, not a measured performance improvement.

Common misconceptions and mistakes

Do not use the test as evidence that group means are equal. Do not describe its usual chi-square calibration as exact for every sample size. Independence and the scientific meaning of comparable spread still matter. Repeated measurements from the same unit require an analysis that respects dependence. Ties, very small groups and degenerate data can affect behavior; if all scores are identical, the statistic's denominator vanishes. The phrase distribution-free should not be expanded into a claim of validity for every pair of populations sharing only the same numerical variance.

Business use

Businesses can use the test to investigate consistency, but should pair it with distribution plots and operationally meaningful spread summaries. Keep measurement units and collection conditions comparable. Identify whether apparent variation comes from process differences, changing case mix or measurement problems. Report the selected centering method and acknowledge small-sample limitations. If many settings or outcomes are tested, account for the resulting multiplicity. A spread test can inform an investigation; it cannot by itself identify the cause of variability or choose a corrective action.

Check your understanding

Three groups have the same median but visibly different dispersion. Is Fligner-Killeen conceptually relevant?

Answer: Yes, for an appropriate independent-group spread comparison. Equal medians do not imply equal spread, and calibration and distributional shape still need attention.

Related terms

Levene test
Homogeneity of variance
Bartlett test
rank transformation
absolute deviation
heteroscedasticity
chi-square approximation
scale parameter

Learn next

Study robust measures of spread and the assumptions behind variance comparisons. Then compare median-centered Levene and Fligner-Killeen procedures, and examine suitable small-sample calibration and models that accommodate unequal variances directly.

Sources

SciPy, fligner reference: https://docs.scipy.org/doc/scipy/reference/generated/scipy.stats.fligner.html SciPy, Fligner-Killeen tutorial: https://docs.scipy.org/doc/scipy/tutorial/stats/hypothesis_fligner.html SciPy implementation, normal-score and test-statistic calculation: https://raw.githubusercontent.com/scipy/scipy/v1.18.0/scipy/stats/_morestats.py

Previous: Flattening transformation Next: Flow-based clustering