Why are there two degrees of freedom?

The numerator and denominator each have their own variation estimate.

Is the mean always one?

No; when it exists it depends on denominator degrees of freedom.

Can the distribution assess a ratio of arbitrary business metrics?

Not without a statistical argument establishing the required distributional structure.

AI Glossary / Statistics and probability / Intermediate

F distribution

Definition

The F distribution is a continuous probability distribution obtained by taking the ratio of two independent chi-square random variables, each divided by its own degrees of freedom. It has two positive degrees-of-freedom parameters, one for the numerator and one for the denominator. It commonly provides reference probabilities for analysis-of-variance tests, regression comparisons, and tests involving population variances.

What it means

Many statistical procedures compare two quantities that estimate variation. When their scaled components have the appropriate independent chi-square distributions, their ratio follows an F distribution. The reference distribution tells us how large or small a ratio can be through random variation under specified assumptions.
An F statistic is not automatically an F-distributed quantity merely because it is a ratio. The construction and assumptions of the statistical procedure are what justify that reference distribution.

How it works

Let U have a chi-square distribution with d_1 degrees of freedom and V have a chi-square distribution with d_2 degrees of freedom, independently. Then F = (U/d_1)/(V/d_2) follows the central F distribution with degrees of freedom (d_1,d_2).
The distribution is supported on positive values and is generally asymmetric, with a right tail. Its shape depends on both degrees of freedom. Reversing their order changes the distribution: if F follows F(d_1,d_2), then 1/F follows F(d_2,d_1).
For independent normal samples, the sample variance divided by the population variance, multiplied by its degrees of freedom, has a chi-square distribution. Under equal population variances, their sample-variance ratio therefore follows an F distribution. Normality and independence are important to that exact result.
In a classical linear-model comparison, an F statistic compares improvement in residual sum of squares per added parameter with an estimate of residual variance. Under the relevant null and normal, independent, constant-variance errors, the numerator and denominator have the required distributions. The details differ from simply comparing two raw sample variances.
Use a cumulative distribution function for lower-tail probabilities, a survival function for upper-tail probabilities, and quantiles for critical values. Software conventions matter: a 95th percentile is a lower-tail cumulative probability of 0.95, whereas some printed tables label that same threshold by an upper-tail probability of 0.05.
The mean is d_2/(d_2 − 2) when d_2 > 2, not exactly one. The variance is finite only when d_2 > 4. Heavy tails at small denominator degrees of freedom can make unusually large ratios quite plausible.
Distinguish the central distribution used under many null hypotheses from a noncentral F distribution used under certain alternatives and in power calculations. A software location shift is not a substitute for noncentrality.

Three practical examples

An analyst uses an F reference distribution for a classical one-way ANOVA comparing group means.
A regression team tests whether a prespecified set of additional coefficients improves a nested normal linear model.
A quality engineer compares two population variances using independent normal samples and the corresponding F distribution.

Think of it this way

Think of comparing two noisy yardsticks of variation. Their ratio may exceed one even when they measure the same underlying variability; the degrees of freedom determine how much ratio fluctuation is expected.

Worked example

For an invented distributional example, let U and V be independent chi-square variables with two degrees of freedom each. Suppose their realized values are U = 6 and V = 2. The ratio is (6/2)/(2/2) = 3, evaluated against F(2,2).
For this special case, the cumulative distribution function is G(x) = x/(1 + x) for x > 0, and the upper-tail probability is 1/(1 + x). Thus P(F ≥ 3) = 1/4 = 0.25.
The 95th percentile solves x/(1 + x) = 0.95, giving x = 19. A ratio of three is therefore not beyond the upper 5% threshold in this example.
This distribution has no finite mean because its denominator degrees of freedom equal two. The example demonstrates why treating all F ratios as tightly concentrated around one is misleading.

Illustrative case study

Fictional case: Alder Analytics compares nested regression models and initially reads a critical-value table with the two degrees of freedom reversed. A review checks which component belongs in the numerator, verifies the software's tail convention, and recalculates the result. The team also examines model assumptions before interpreting the test. This invented case illustrates that a correct statistic still needs the correct reference distribution.

Common misconceptions and mistakes

Do not assume an F statistic can be negative or that the central F distribution is symmetric around one. Do not swap numerator and denominator degrees of freedom without taking the reciprocal statistic.
A large F value is evidence against a specified null only when the procedure's reference assumptions hold. It is not an effect size or proof of practical importance.
An ANOVA F test concerns mean structure, despite being constructed from quantities called mean squares.

Business use

Understanding the F distribution helps teams interpret common regression and process-analysis outputs. Reports should identify the tested hypothesis, statistic, both degrees of freedom, and relevant assumptions. Business decisions also need effect estimates, uncertainty, and operational importance; a tail probability alone cannot provide those.

Check your understanding

What condition connects U and V in the definition?

Answer: Independence.

What is the upper-tail probability at three for F(2,2)?

Answer: 0.25.

Why is 19 the 95th percentile in the example?

Answer: Its cumulative probability is 19/20 = 0.95.

Related terms

Degrees of freedom
Critical value
Survival function
Chi-square distribution
ANOVA
variance ratio
nested model
noncentral F distribution

Learn next

Learn chi-square distributions and normal-sample variance theory. Then derive a classical ANOVA F statistic and distinguish its null distribution from its alternative distribution.

Sources

SciPy, scipy.stats.f documentation: https://docs.scipy.org/doc/scipy/reference/generated/scipy.stats.f.html NIST/SEMATECH, “F Distribution”: https://www.itl.nist.gov/div898/handbook/eda/section3/eda3665.htm

Previous: F 1 Next: f-divergence