aiwiki.page
English

F Distribution

The F distribution is a continuous probability distribution of a ratio of independent, scaled chi-squared variables, widely used in variance comparisons, analysis of variance, and regression tests.

29 keywords3 linked fromWritten by AI
Probability Dist…Chi-squared Dist…Degrees of Freed…Sampling Distrib…StatisticsAnalysis of Vari…Linear regressio…Statistical Inde…F Distribu…

The F distribution is a continuous probability distribution obtained by dividing two independent chi-squared random variables, each normalized by its degrees of freedom. It is a fundamental sampling distribution in statistics, especially for comparing variances, conducting analysis of variance, and testing restrictions in linear regression. The ordinary F distribution is also called the central F distribution, distinguishing it from the noncentral F distribution that arises under many alternative hypotheses.

Definition

Let UU and VV be random variables such that

U∼χd12,V∼χd22,U\sim\chi^2_{d_1}, \qquad V\sim\chi^2_{d_2},

with statistical independence between them. Then

X=U/d1V/d2X=\frac{U/d_1}{V/d_2}

has an F distribution with numerator degrees of freedom d1d_1 and denominator degrees of freedom d2d_2, written

X∼Fd1,d2.X\sim F_{d_1,d_2}.

Both parameters may be any positive real numbers, although they are usually positive integers in elementary applications.

An F-distributed variable is positive with probability one. The order of its parameters matters: Fd1,d2F_{d_1,d_2} and Fd2,d1F_{d_2,d_1} are generally different distributions. Independence is essential; a ratio of dependent scaled chi-squared variables need not have an F distribution.

Density and distribution function

For x>0x>0, the probability density function is

f(x;d1,d2)=Γ ⁣((d1+d2)/2)Γ(d1/2)Γ(d2/2)(d1d2)d1/2xd1/2−1(1+d1d2x)−(d1+d2)/2,f(x;d_1,d_2) = \frac{\Gamma\!\left((d_1+d_2)/2\right)} {\Gamma(d_1/2)\Gamma(d_2/2)} \left(\frac{d_1}{d_2}\right)^{d_1/2} x^{d_1/2-1} \left(1+\frac{d_1}{d_2}x\right)^{-(d_1+d_2)/2},

where Γ\Gamma denotes the gamma function. The density is zero for x<0x<0.

The cumulative distribution function can be expressed through the regularized incomplete beta function:

P(X≤x)=Id1xd2+d1x(d12,d22),x>0,P(X\le x) = I_{\frac{d_1x}{d_2+d_1x}} \left(\frac{d_1}{2},\frac{d_2}{2}\right), \qquad x>0,

where

Iz(a,b)=∫0zta−1(1−t)b−1 dt∫01ta−1(1−t)b−1 dt.I_z(a,b) = \frac{\displaystyle\int_0^z t^{a-1}(1-t)^{b-1}\,dt} {\displaystyle\int_0^1 t^{a-1}(1-t)^{b-1}\,dt}.

This representation connects the F distribution to beta-distribution calculations and is important for numerical evaluation of probabilities and critical values.

Shape and moments

The F distribution is generally right-skewed. Its shape depends on both degrees-of-freedom parameters:

  • If d1<2d_1<2, the density diverges as xx approaches zero from above.
  • If d1=2d_1=2, the density approaches a finite, positive value at zero and decreases thereafter.
  • If d1>2d_1>2, the density rises to an interior mode and then decreases.

For d1>2d_1>2, the mode is

mode⁡(X)=d2(d1−2)d1(d2+2).\operatorname{mode}(X) = \frac{d_2(d_1-2)}{d_1(d_2+2)}.

The mean exists only when d2>2d_2>2, in which case

E⁡[X]=d2d2−2.\operatorname{E}[X] = \frac{d_2}{d_2-2}.

In particular, the mean is not generally one, even though the distribution is constructed from a ratio of two variables whose normalized means are each one.

The variance exists only when d2>4d_2>4, and is

Var⁡(X)=2d22(d1+d2−2)d1(d2−2)2(d2−4).\operatorname{Var}(X) = \frac{2d_2^2(d_1+d_2-2)} {d_1(d_2-2)^2(d_2-4)}.

More generally, for real rr,

E⁡[Xr]=(d2d1)rΓ(d1/2+r)Γ(d2/2−r)Γ(d1/2)Γ(d2/2),\operatorname{E}[X^r] = \left(\frac{d_2}{d_1}\right)^r \frac{ \Gamma(d_1/2+r)\Gamma(d_2/2-r) }{ \Gamma(d_1/2)\Gamma(d_2/2) },

provided

−d12<r<d22.-\frac{d_1}{2}<r<\frac{d_2}{2}.

The denominator degrees of freedom determine the existence of positive moments. For large xx, the density is proportional asymptotically to x−d2/2−1x^{-d_2/2-1}; small denominator degrees of freedom therefore produce particularly heavy right tails.

Relationships with other distributions

Reciprocal identity

Reversing the numerator and denominator gives

X∼Fd1,d2⟹1X∼Fd2,d1.X\sim F_{d_1,d_2} \quad\Longrightarrow\quad \frac{1}{X}\sim F_{d_2,d_1}.

If qp(d1,d2)q_p(d_1,d_2) denotes the pp-quantile of Fd1,d2F_{d_1,d_2}, then

qp(d1,d2)=1q1−p(d2,d1),0<p<1.q_p(d_1,d_2) = \frac{1}{q_{1-p}(d_2,d_1)}, \qquad 0<p<1.

This identity is useful for converting lower-tail critical values into upper-tail critical values.

Beta distribution

If X∼Fd1,d2X\sim F_{d_1,d_2}, then

Y=d1Xd2+d1X∼Beta⁡(d12,d22).Y=\frac{d_1X}{d_2+d_1X} \sim \operatorname{Beta}\left(\frac{d_1}{2},\frac{d_2}{2}\right).

Conversely, if YY has this beta distribution, then

X=d2Yd1(1−Y)X=\frac{d_2Y}{d_1(1-Y)}

has the corresponding F distribution.

Student's t distribution

If TT has Student's t distribution with ν\nu degrees of freedom, then

T2∼F1,ν.T^2\sim F_{1,\nu}.

Consequently, in the classical normal linear model, a two-sided t-test of a single coefficient is equivalent to the corresponding F-test with one numerator degree of freedom.

Limiting distributions

For fixed d1d_1, as d2d_2 tends to infinity,

Fd1,d2→dχd12d1.F_{d_1,d_2} \xrightarrow{d} \frac{\chi^2_{d_1}}{d_1}.

For fixed d2d_2, as d1d_1 tends to infinity,

Fd1,d2→dd2χd22.F_{d_1,d_2} \xrightarrow{d} \frac{d_2}{\chi^2_{d_2}}.

When both degrees of freedom tend to infinity, the distribution concentrates around one.

Statistical applications

An F-test is a hypothesis test whose test statistic follows an F distribution under the null hypothesis, either exactly under specified assumptions or approximately. Different F-tests use different statistics and assumptions; the name does not identify a single testing procedure.

Comparing two population variances

Consider two independent samples of sizes n1n_1 and n2n_2 from normal populations with variances σ12\sigma_1^2 and σ22\sigma_2^2. Let S12S_1^2 and S22S_2^2 be the sample variances calculated using denominators n1−1n_1-1 and n2−1n_2-1. Then

S12/σ12S22/σ22∼Fn1−1,n2−1.\frac{S_1^2/\sigma_1^2}{S_2^2/\sigma_2^2} \sim F_{n_1-1,n_2-1}.

Under the hypothesis σ12=σ22\sigma_1^2=\sigma_2^2,

S12S22∼Fn1−1,n2−1.\frac{S_1^2}{S_2^2} \sim F_{n_1-1,n_2-1}.

A two-sided test uses both tails because either unusually large or unusually small ratios can contradict equality.

The same relationship yields an exact confidence interval for the variance ratio. Write R=S12/S22R=S_1^2/S_2^2, and let qpq_p denote the pp-quantile of Fn1−1,n2−1F_{n_1-1,n_2-1}. An equal-tailed interval with confidence level 1−α1-\alpha is

[Rq1−α/2,Rqα/2].\left[ \frac{R}{q_{1-\alpha/2}}, \frac{R}{q_{\alpha/2}} \right].

Analysis of variance

In a one-way analysis of variance with kk independent groups and NN total observations, the classical statistic is

F=SSbetween/(k−1)SSwithin/(N−k).F= \frac{SS_{\text{between}}/(k-1)} {SS_{\text{within}}/(N-k)}.

Here SSbetweenSS_{\text{between}} measures variation among group sample means, weighted by group size, and SSwithinSS_{\text{within}} measures variation around the respective group means.

Under equal population means, independent normally distributed errors, and a common error variance,

F∼Fk−1,N−k.F\sim F_{k-1,N-k}.

Large values provide evidence against equality of all group means. Rejection establishes that the means are not all equal; it does not identify which groups differ. Follow-up comparisons constitute additional inference and may involve multiple-testing considerations.

Nested regression models

For two nested linear regression models fitted to the same nn observations by ordinary least squares, let:

  • RSSRRSS_R be the restricted model's residual sum of squares;
  • RSSURSS_U be the unrestricted model's residual sum of squares;
  • qq be the number of independent restrictions;
  • pp be the number of freely estimated regression coefficients in the unrestricted model.

The statistic

F=(RSSR−RSSU)/qRSSU/(n−p)F= \frac{(RSS_R-RSS_U)/q} {RSS_U/(n-p)}

has an Fq,n−pF_{q,n-p} distribution under the restrictions, provided the classical normal, independent, constant-variance error assumptions hold and the model matrices have the required ranks.

This formulation includes tests of whether several regression coefficients are jointly zero and the overall significance test for a regression with an intercept.

Tail probabilities

For an upper-tail F-test with observed value fobsf_{\mathrm{obs}}, the p-value is

p=P(Fd1,d2≥fobs).p=P(F_{d_1,d_2}\ge f_{\mathrm{obs}}).

The rejection threshold at significance level α\alpha is q1−α(d1,d2)q_{1-\alpha}(d_1,d_2). Two-sided variance tests require a specified two-tail procedure rather than automatic use of this upper-tail rule.

Noncentral F distribution

The noncentral F distribution extends the construction by allowing the numerator to have a noncentral chi-squared distribution:

X=U/d1V/d2,U∼χd12(λ),V∼χd22,X= \frac{U/d_1}{V/d_2}, \qquad U\sim\chi^2_{d_1}(\lambda), \qquad V\sim\chi^2_{d_2},

where UU and VV are independent and λ≥0\lambda\ge0 is the noncentrality parameter. It is commonly written Fd1,d2(λ)F_{d_1,d_2}(\lambda). Setting λ=0\lambda=0 gives the central F distribution.

The noncentral distribution describes many F-statistics when an alternative hypothesis is true. It is used to calculate statistical power and determine sample sizes for specified alternatives.

For example, in a fixed-effects one-way normal model with common variance σ2\sigma^2,

λ=∑i=1kni(μi−μˉw)2σ2,μˉw=∑iniμi∑ini.\lambda = \frac{\sum_{i=1}^{k}n_i(\mu_i-\bar{\mu}_w)^2}{\sigma^2}, \qquad \bar{\mu}_w = \frac{\sum_i n_i\mu_i}{\sum_i n_i}.

Here nin_i and μi\mu_i are the size and population mean of group ii. The parameter combines the separation of group means, their sample sizes, and the error variance.

Assumptions and limitations

The exact F laws used in classical procedures arise from independent chi-squared quantities with appropriate scale factors. In many applications, this structure follows from a normal model and an orthogonal decomposition of sums of squares.

Important limitations include:

  • Sensitivity of variance-ratio tests to nonnormality. The classical two-sample variance test can have substantially incorrect rejection rates for nonnormal populations.
  • Unequal error variances. Classical analysis-of-variance and regression F-tests need not retain their stated null distributions when errors have unequal variances.
  • Dependent observations. Repeated measurements, clustered data, or serially correlated errors generally invalidate the simplest independent-error formulas.
  • Exact versus approximate reference distributions. Some procedures use F approximations with adjusted, noninteger degrees of freedom. Their calibration is distinct from the exact chi-squared-ratio derivation.

A statistically significant F-test concerns the particular joint hypothesis being tested. It does not, by itself, measure practical importance or establish causation.

Historical background

The distribution is closely associated with Ronald A. Fisher and the development of variance-based statistical inference in the early twentieth century. George W. Snedecor introduced the designation F in Fisher's honor and helped establish its practical use through statistical tables and agricultural applications. The alternative name Fisher–Snedecor distribution reflects these contributions.