Avoiding Degrees-of-Freedom and Tail-Reversal Errors

Why F-Test Bookkeeping Matters

An F test can be set up with the right formula and still give the wrong conclusion if the sample labels, degrees of freedom, or tail are mismatched. The statistic is a ratio, so reversing its numerator and denominator changes both its value and its reference distribution. The direction of the alternative determines which tail supplies the rejection region.

This tutorial focuses on those bookkeeping errors. The exact classical test assumes two independent random samples from normal populations. If the samples are dependent or the populations are not normal, the usual exact F-test calibration need not apply; those are separate limitations from the arithmetic errors considered here.

Definition: Let two independent samples from normal populations have sizes $n_1$ and $n_2$, sample variances $S_1^2$ and $S_2^2$, and population variances $\sigma_1^2$ and $\sigma_2^2$. Under the null hypothesis $H_0:\sigma_1^2=\sigma_2^2$, $$ \frac{S_1^2}{S_2^2}\sim F(n_1-1,n_2-1). $$ The first degrees-of-freedom parameter belongs to the sample variance in the numerator; the second belongs to the sample variance in the denominator.

The sample variance uses $n-1$ degrees of freedom because the sample mean is estimated from the same $n$ observations. Thus, a sample of size 9 contributes 8 degrees of freedom—not 9—when its variance is used in the classical F ratio.

Ratio Order, Degrees of Freedom, and Tails

For clarity, write the ratio in the same order as the population variances named in the alternative. If the alternative is $\sigma_1^2>\sigma_2^2$, use $S_1^2/S_2^2$ and an upper-tail test. If the alternative is $\sigma_1^2<\sigma_2^2$, that same ratio calls for a lower-tail test. If the alternative is that the variances differ, the test is two-sided.

Define $c_\alpha(a,b)$ as the upper-tail critical value for an $F(a,b)$ distribution:

P(F(a,b)>c_\alpha(a,b))=\alpha.

A lower-tail critical value is not generally the same table entry with the same degrees of freedom. The reciprocal identity explains the correct conversion.

Theorem: If $F\sim F(a,b)$, then $1/F\sim F(b,a)$. Consequently, the lower-tail $\alpha$ critical value for $F(a,b)$ is $1/c_\alpha(b,a)$. Reversing a sample-variance ratio therefore requires swapping the degrees of freedom as well.

Here is a short justification. If $U$ and $V$ are independent chi-square random variables with $a$ and $b$ degrees of freedom, respectively, then $F=(U/a)/(V/b)$ has an $F(a,b)$ distribution. Taking the reciprocal gives

\frac{1}{F}=\frac{V/b}{U/a}\sim F(b,a),

because the numerator and denominator have exchanged roles. In particular, for $F\sim F(a,b)$,

P(F<1/c_\alpha(b,a)) =P(1/F>c_\alpha(b,a)) =\alpha.

This identity is useful for interpreting lower-tail tests with tables that provide upper-tail critical values. It does not mean that one may reverse a ratio while keeping its degrees of freedom or its tail unchanged.

Worked Examples

Worked Example: A One-Sided Test With the Samples Reversed

Two independent normal samples are taken to compare hypothetical claim-severity variances for two coverage groups. Group A has $n_A=9$ observations and sample variance $S_A^2=36$; group B has $n_B=16$ observations and sample variance $S_B^2=12$. Test $H_0:\sigma_A^2=\sigma_B^2$ against $H_1:\sigma_A^2>\sigma_B^2$ at $\alpha=0.05$. Use the supplied upper-tail table value $c_{0.05}(8,15)=2.64$.

Method. Put the variance from the population named as larger under the alternative in the numerator. The observed statistic is

F_{\text{obs}}=\frac{S_A^2}{S_B^2} =\frac{36}{12}=3.

The numerator sample has $9-1=8$ degrees of freedom, and the denominator sample has $16-1=15$. Thus the reference distribution under the null is $F(8,15)$. This is an upper-tail test because $H_1$ says that A has the larger variance. Since $3>2.64$, reject $H_0$ at the 5% level.

Check by reversing the ratio correctly. If instead the statistic is written as $S_B^2/S_A^2$, its observed value is $12/36=1/3$, and its degrees of freedom are $(15,8)$. It is now a lower-tail test. The corresponding lower critical value is $1/2.64\approx0.3788$, so $1/3<0.3788$ gives the same rejection.

Common mistake avoided: Using $S_B^2/S_A^2=1/3$ but comparing it with the upper-tail $F(8,15)$ critical value 2.64 mixes a reversed ratio with the original degrees of freedom and tail. The correct reversed setup uses $F(15,8)$ and the lower tail.

Worked Example: A Two-Sided Test and the Tail Allocation

Two independent normal samples have sizes $n_1=12$ and $n_2=10$, with sample variances $S_1^2=25$ and $S_2^2=9$. Test whether the population variances differ at $\alpha=0.10$. For this exercise, use the supplied rounded table values $c_{0.05}(11,9)=3.20$ and $c_{0.05}(9,11)=3.00$.

Method. The alternative is two-sided: $H_1:\sigma_1^2\ne\sigma_2^2$. Use the ratio in the stated sample order:

F_{\text{obs}}=\frac{25}{9}=2.7778\ldots, \qquad (\nu_1,\nu_2)=(12-1,10-1)=(11,9).

A 10% two-sided test places $0.10/2=0.05$ in each tail. The upper boundary is $c_{0.05}(11,9)=3.20$. For the lower boundary, use the reciprocal rule and swap the degrees of freedom:

\text{lower boundary} =\frac{1}{c_{0.05}(9,11)} =\frac{1}{3.00} =0.3333\ldots.

The observed value satisfies $0.3333<2.7778<3.20$, so it is not in either rejection region. Do not reject $H_0$ at the 10% level.

Common mistake avoided: Using an upper-tail 10% critical value as the upper boundary would fail to reserve 5% for the lower tail. A two-sided test at level 0.10 uses 0.05 in each tail, not 0.10 in one tail.

Worked Example: A Lower-Tail Test and a Degrees-of-Freedom Check

Two independent normal samples measure hypothetical repair costs under two contract designs. Design 1 has $n_1=8$ observations and sample variance $S_1^2=1$; design 2 has $n_2=13$ and sample variance $S_2^2=4$. Test $H_0:\sigma_1^2=\sigma_2^2$ against $H_1:\sigma_1^2<\sigma_2^2$ at $\alpha=0.05$. The upper-tail table gives $c_{0.05}(12,7)=3.57$.

Method. Keep the sample order aligned with the alternative. The observed statistic and degrees of freedom are

F_{\text{obs}}=\frac{S_1^2}{S_2^2}=\frac{1}{4}=0.25, \qquad (\nu_1,\nu_2)=(8-1,13-1)=(7,12).

The alternative is lower-tailed. Convert the supplied upper-tail value with the degrees of freedom swapped:

\text{lower critical value for }F(7,12) =\frac{1}{c_{0.05}(12,7)} =\frac{1}{3.57} \approx0.2801.

Because $0.25<0.2801$, reject $H_0$. The sample evidence supports a smaller variance for design 1 at the 5% level.

Common mistake avoided: The upper-tail table entry $3.57$ is not the lower-tail cutoff for $F(7,12)$. Also, using $8$ and $13$ rather than $7$ and $12$ would assign the wrong degrees of freedom to the ratio.

A Reliable Diagnostic Sequence

Before comparing an observed statistic with a critical value, check the entire setup as a connected chain. Each item depends on the choices made before it.

State the direction of the alternative.
Identify whether the claim is that population 1 has the larger variance, the smaller variance, or simply a different variance.

Write the ratio explicitly.
Record which sample variance is the numerator and which is the denominator. Do not rely on an unstated convention such as “larger variance on top.”

Attach the degrees of freedom to their samples.
The numerator contributes $n_{\text{numerator}}-1$; the denominator contributes $n_{\text{denominator}}-1$.

Match the critical-value tail.
A larger-variance alternative in the numerator calls for an upper tail; a smaller-variance alternative calls for a lower tail. A two-sided test divides its significance level between both tails.

Check any reversal before deciding.
If the ratio is inverted, swap the degrees of freedom and use the reciprocal-tail relationship before making the comparison.

These checks matter because an F distribution is generally asymmetric, and its shape depends on both degrees of freedom. A ratio and a critical value that each look numerically plausible can still be incompatible. The most dependable check is to say, in words, what a large observed ratio means for the alternative, then verify that the rejection region is in that direction.

Key takeaway: The numerator determines the first degrees of freedom, the denominator determines the second, and the alternative determines the tail. If you reverse the ratio, reverse the degrees of freedom and convert the tail as well.

Check Your Understanding

For each problem, treat the samples as independent and normally distributed. Critical values stated in the question are supplied lookup values; use them as given.

Samples A and B have sizes 11 and 6, with sample variances 18 and 6. Test $H_1:\sigma_A^2>\sigma_B^2$ at 5%. The supplied upper critical value is $c_{0.05}(10,5)=4.50$. State the ratio, degrees of freedom, and decision. (Answer: $F_{\text{obs}}=3$, degrees of freedom $(10,5)$; do not reject $H_0$.)
Samples C and D have sizes 7 and 15, with sample variances 2 and 8. Test $H_1:\sigma_C^2<\sigma_D^2$ at 5%. The supplied value is $c_{0.05}(14,6)=3.50$. Find the lower critical value and decision. (Answer: $F_{\text{obs}}=0.25$, degrees of freedom $(6,14)$, lower critical value $1/3.50\approx0.2857$; reject $H_0$.)
For a two-sided test at $\alpha=0.02$, samples of sizes 9 and 14 have sample variances 16 and 4. The supplied values are $c_{0.01}(8,13)=4.10$ and $c_{0.01}(13,8)=4.50$. Find both rejection boundaries and decide. (Answer: $F_{\text{obs}}=4$, degrees of freedom $(8,13)$, boundaries $1/4.50\approx0.2222$ and $4.10$; do not reject $H_0$.)
Samples of sizes 10 and 18 have sample variances 5 and 20. For $H_1:\sigma_1^2<\sigma_2^2$ at 5%, the supplied value is $c_{0.05}(17,9)=3.00$. Write the ratio in the stated order and decide. (Answer: $F_{\text{obs}}=0.25$, degrees of freedom $(9,17)$, lower critical value $1/3\approx0.3333$; reject $H_0$.)
For a two-sided test at $\alpha=0.10$, samples of sizes 8 and 12 have sample variances 6 and 3. The supplied values are $c_{0.05}(7,11)=3.50$ and $c_{0.05}(11,7)=3.80$. Find the boundaries and decision. (Answer: $F_{\text{obs}}=2$, degrees of freedom $(7,11)$, boundaries $1/3.80\approx0.2632$ and $3.50$; do not reject $H_0$.)

What You'll Learn