Is this the same as ordinary Levene's test?

It is a related median-centered variant; the original version centers on the mean.

Can group sizes differ?

Yes, although design and small-sample behavior still matter.

Does it prove two services are equally reliable?

No. Reliability involves more than one spread comparison.

AI Glossary / Statistics and probability / Advanced

Brown - Forsythe test

Definition

The Brown - Forsythe test commonly means a median-centered version of Levene's test for equality of population variances across groups. It compares absolute distances from each group's median. The name also appears in a different procedure for comparing means when variances are unequal, so the intended hypothesis must be stated.

What it means

This entry concerns the variance test. Two services can have similar typical response times yet differ in consistency. One produces tightly clustered times; the other has much wider variation. Testing spread addresses that distinction rather than asking which service is faster on average.
The median provides a center that is less affected by an extreme observation than the mean. However, absolute distances from that center still reflect extreme values. Calling the procedure robust does not mean that unusual observations, dependence, or poor sampling can be ignored.

How it works

For each independent group, calculate its sample median. Replace every observation with its absolute distance from that median. Apply the usual one-way analysis-of-variance calculation to these distances, comparing variation between their group means with variation within groups.
With k groups and N total observations, the statistic is the between-group sum of squares divided by k minus one, over the within-group sum of squares divided by N minus k. Its conventional reference is an approximate F distribution with those degrees of freedom.
The stated null is equality of population variances. Interpretation is clearest when groups have comparable distributional shapes: the calculation directly compares average absolute deviations, which are not identical to variances for arbitrary distributions. Inspect the data alongside the test.

Three practical examples

An AI service team compares response-time variability across independent operating configurations. It checks that repeated calls from the same session are not incorrectly treated as independent evidence.
A quality analyst investigates consistency across production lines. A line with a satisfactory average measurement may still produce an unacceptably broad range of measurements.
A model evaluator compares error spreads across separately sampled groups. The analyst reports group sizes and descriptive spread measures rather than presenting a p-value as a complete account of performance.

Think of it this way

Picture two groups of people standing around separate meeting points. First measure how far each person stands from their own group's center. Then compare those distances. Moving an entire group to a different meeting point does not change how dispersed its members are around that point.

Worked example

Consider two invented groups of four measurements. Group A contains 8, 9, 11, and 12. Group B contains 4, 8, 12, and 16. Both medians equal 10.
For A, the absolute distances from 10 are 2, 1, 1, and 2. Their average is 1.5. For B, the distances are 6, 2, 2, and 6, averaging 4. The overall average distance across all eight observations is 2.75.
The between-group sum of squares is 4 times (1.5 minus 2.75) squared, plus 4 times (4 minus 2.75) squared. This equals 12.5.
Within A, the four squared differences from its distance-average of 1.5 sum to 1. Within B, the squared differences from 4 sum to 16. The combined within-group sum of squares is therefore 17.
There are two groups and eight observations. The numerator degrees of freedom are one, and the denominator degrees of freedom are six. The statistic is (12.5/1)/(17/6), or 75/17, approximately 4.412. A conventional analysis compares this with an F distribution having one and six degrees of freedom. The tiny dataset illustrates the calculation, not dependable small-sample calibration.
The ordinary sample variances of the original measurements are 10/3 and 80/3. Their ratio is eight, but eight is not the Brown - Forsythe statistic. The method analyzes absolute deviations and their variability, rather than simply dividing the original sample variances.
Now add 100 to every B observation. Its median becomes 110, while its four absolute deviations remain 6, 2, 2, and 6. The test statistic stays unchanged even though the group means become very different. This makes the distinction between location and spread concrete.
If both groups are instead multiplied by ten, both sums of squares for the transformed distances increase by a factor of 100. Their ratio, and therefore the test statistic, again stays unchanged.

Illustrative case study

Fictional company Cedar Service compares two independently sampled operating configurations. Its dashboard initially shows only average response time. An analyst adds medians, spread summaries, and distribution plots, then uses a prespecified median-centered variance comparison.
The team investigates whether workload composition differs between configurations before attributing variability to the configuration itself. It also checks whether several measurements came from the same session. If so, the sampling structure requires appropriate treatment rather than counting every row as independent.
The business decision considers operationally troublesome delays as well as the statistical comparison. A narrower distribution is useful only in relation to the service objective. This fictional scenario illustrates analysis planning, not a documented deployment or promised improvement.

Common misconceptions and mistakes

The variance test does not test whether group medians are equal. Medians are used to construct the deviations.
It is also distinct from Brown - Forsythe procedures for comparing means under unequal variances. A software option bearing the name needs its documentation checked.
A nonsignificant result does not establish equal variances. Limited data may leave substantial differences uncertain.
A significant result identifies evidence against the overall null, not automatically which specific pairs differ or why.
Robustness to some nonnormal distributions does not remove the requirement to consider independence and distributional differences.

Business use

Use the test as one part of a consistency assessment. Define the sampling unit, groups, outcome, and practical tolerance before analysis. Report descriptive spread alongside the test, and distinguish statistical evidence from the operational size of a difference. Investigate sampling or workload explanations before claiming a causal effect.

Check your understanding

What is compared in the example?

Answer: Absolute distances from each group's median.

Is eight the test statistic?

Answer: No; eight is the original sample-variance ratio.

Why does adding 100 to B change nothing?

Answer: Its deviations from its own median remain identical.

Related terms

Learn next

Compare median-centered Levene testing with Bartlett and Fligner - Killeen tests. Study how unequal-variance mean comparisons answer a separate question.

Sources

NIST/SEMATECH, Levene Test for Equality of Variances: https://www.itl.nist.gov/div898/handbook/eda/section3/eda35a.htm . SciPy, levene: https://docs.scipy.org/doc/scipy/reference/generated/scipy.stats.levene.html . Statsmodels, anova_oneway, for the distinct means procedure: https://www.statsmodels.org/stable/generated/statsmodels.stats.oneway.anova_oneway.html . Numerical data and company scenario are original illustrations.

Previous: Brown - Conrady model Next: Brownout (software engineering)