Statistics › Section K · The chi-square distribution › Topic K.6

Test of a single variance

Machines are trusted for their consistency, not just their aim: a filler must not only average 500 ml but stay tight around it. The single-variance test prices claims about spread - with a chi-square statistic that measures how the sample's variance argues with the claimed one.

Claims about spread

Some claims are about spread, not centre: this filling process stays within a tight tolerance; this investment's volatility is 12% a year; this scale's measurement error is known. The parameter is \(\sigma^2\) (the SD's square, because the algebra runs on variances), and the null takes the usual form \(H_0: \sigma^2 = \sigma_0^2\), with the alternative carrying the question's direction. The test statistic is a chi-square statistic - the family's fourth and final job.

The statistic

The sample variance \(s^2\) argues with the claimed \(\sigma_0^2\) through

\chi^2 = \frac{(n-1)s^2}{\sigma_0^2}

with \(n - 1\) degrees of freedom - the same \(n-1\) that makes the sample variance unbiased. The reading is intuitive: if \(s^2\) matches the claim, the ratio \(s^2/\sigma_0^2\) sits near 1 and \(\chi^2\) lands near \(n-1\), the curve's mean. A sample varying far MORE than claimed inflates \(\chi^2\) into the right tail; far LESS than claimed starves it into the left tail. The statistic IS the ratio of observed to claimed variance, scaled.

Two tails on a skewed curve

Spread can fail in two directions - too noisy or too eerily steady - so the natural alternative \(H_a: \sigma^2 \ne \sigma_0^2\) is two-sided, and both tails of the chi-square curve carry evidence. A large \(\chi^2\): the process is noisier than claimed. A small one: the process is quieter than claimed (or the data are too good to be true: oddly small variation can flag tidied or fabricated data, the charge R. A. Fisher famously, and still debatably, raised about Mendel's pea counts). The curve's right skew breaks the tail symmetry the z-tests enjoyed: the two cutoffs sit at different distances from the centre, so a two-sided p-value is found by doubling the area of the tail the statistic actually lands in.

Worked example: the filler's consistency

A filling line claims \(\sigma = 2\) ml. A quality check measures \(n = 25\) fills and finds \(s = 3\) ml. Is the process noisier than claimed? 1. \(H_0: \sigma^2 = 4\); \(H_a: \sigma^2 > 4\) - one-sided, the question names a direction. 2. \(\chi^2 = \frac{24 \times 9}{4} = 54\) with df = 24. 3. The df = 24 curve has mean 24 and SD \(\sqrt{48} \approx 6.9\): 54 sits over four SDs above centre. 4. \(p \approx 0.0004 \le 0.05\): reject overwhelmingly - the process is far noisier than the 2 ml claim, and the investigation turns to the machine.

Worked example: the scale's stability

A lab claims its scale's measurement error has \(\sigma = 0.5\) mg. Twenty repeated weighings of a standard give \(s = 0.30\) mg - steadier than claimed. 1. \(H_0: \sigma^2 = 0.25\); \(H_a: \sigma^2 \ne 0.25\) - two-sided: either direction indicts the claim. 2. \(\chi^2 = \frac{19 \times 0.09}{0.25} = 6.84\) with df = 19. 3. The df = 19 curve has mean 19: 6.84 lies far in the LEFT tail - the data are too quiet for the claim. 4. The left-tail area \(P(\chi^2 < 6.84) \approx 0.005\); doubled for the two-sided question, \(p \approx 0.010 \le 0.05\): reject - the claim's error spread is wrong, in the flattering direction. Check: the df-19 curve's lower 2.5% cutoff is about 8.91, and 6.84 sits below it, agreeing.

Reading the verdict in SD units

The test's conclusion lands best back in the original units. Rejecting \(\sigma^2 = 4\) says the claimed SD of 2 ml is wrong - and the sample's \(s = 3\) ml says it is wrong by half again: fills scatter about 50% more widely than the specification promised, a difference a process engineer can act on. The scale's rejection is the mirror image: the true error SD is plausibly nearer 0.3 mg than the claimed 0.5 - good news, but news all the same, since downstream calculations built on 0.5 are now over-conservative. Variances are the algebra's language; SDs are the workshop's.

Conditions: normality matters here

The variance test is the family's most demanding customer: it assumes the population itself is normal, and it is NOT robust to that assumption failing. The t-tests forgave moderate skew via the central limit theorem - averages behave - but \(s^2\) is built from squared deviations, and squared values amplify exactly the long tails and outliers that non-normal data carry. A skewed population makes the stated p-value unreliable at any sample size. Plot the data first; if the shape is not plausibly normal, an alternative (a transformation, or a robust spread comparison) is the honest route.

Common mistakes to avoid

First, using \(n\) instead of \(n-1\) in the numerator - the same slip that biases the variance itself. Second, testing \(\sigma\) directly: the algebra wants variances, so square the claimed SD before dividing. Third, reading only the right tail: small \(\chi^2\) is evidence too, of suspicious steadiness. Fourth, trusting the test on visibly non-normal data - this one assumption the test truly needs. Fifth, forgetting that the df is \(n-1\), giving the curve its centre. Sixth, concluding "the process is consistent" from a non-rejection: consistency with the claim is not proof of the claim.

Key takeaways

Claims about spread run through \(\chi^2 = (n-1)s^2/\sigma_0^2\) on \(n-1\) degrees of freedom: near \(n-1\) means the sample matches the claim, the right tail means noisier, the left tail means quieter - and a two-sided test reads both tails of the skewed curve separately. The filler: \(\chi^2 = 54\) at df 24, \(p \approx 0.0004\) - far noisier than the 2 ml claim. The scale: \(\chi^2 = 6.84\) at df 19, left tail \(\approx 0.005\), two-sided \(p \approx 0.010\) - too steady for the claim. The test demands a genuinely normal population, uniquely among this chapter's tests, and squared claims (variances, not SDs) at the door.

Flashcards

Flashcard

χ² = (n−1)s²/σ0²

The single-variance test statistic

Observed over claimed variance, scaled.

Test your understanding

Question 1 of 5

A process claims σ = 2 ml; a sample of 25 gives s = 3 ml. The test statistic is:

Question 2 of 5

The filler's χ² = 54 at df 24. Relative to the curve's centre, this sits:

Question 3 of 5

A chi-square variance statistic deep in the LEFT tail means the sample is:

Question 4 of 5

Unlike the t tests, the single-variance chi-square test requires:

Question 5 of 5

A lab claims σ = 0.5 mg; 20 weighings give s = 0.30 mg. Testing H0: σ² = 0.25 two-sided, the statistic is:

← Previous topicNext topic →