Statistics › Section K · The chi-square distribution › Topic K.1
The chi-square distribution
Means are compared with t and proportions with z - but counts in categories need a new curve. The chi-square distribution prices how far observed counts depart from expected ones, and this topic meets the family before its tests put it to work.
A curve for counts
Means are compared with \(t\), proportions with \(z\) - but questions about counts in categories (are these dice fair? do these two variables travel together?) need a new curve. The chi-square distribution is the reference curve for count-based tests: it prices how far observed counts can reasonably depart from the counts a null hypothesis expects. Three tests in the coming topics - goodness-of-fit, independence, homogeneity - plus a variance test, all run on this one family.
The chi-square idea: squared distances
The statistic that feeds this distribution measures departure:
Each category contributes the squared gap between its observed count and its expected count, scaled by the expected count's size, and the contributions add. Squaring makes every departure count positively - under-filled and over-filled categories contribute alike - and scaling by expected stops big categories from dominating by mere size. Large \(\chi^2\) means large total departure; \(\chi^2\) near zero means the data hug the expectations.
The distribution's shape
The chi-square distribution lives on the non-negative half-line and leans right. Values are never negative - squared gaps cannot be - so the curve lives on zero and above and stretches a long right tail toward large departures. For df of 3 or more it rises from zero to a peak and falls away; at df 1 and 2 it is highest at zero itself. Unlike the symmetric normal, its mean sits just to the right of its peak, and small values are common while very large ones are rare but possible. The curve's whole message: modest departures from expectation are ordinary; enormous ones are worth explaining.
One parameter: degrees of freedom
The family's single parameter is the degrees of freedom, and each test supplies its own: \(k - 1\) for a goodness-of-fit test over \(k\) categories, \((r-1)(c-1)\) for a table with \(r\) rows and \(c\) columns, \(n - 1\) for a single variance - formulas the coming topics justify. The df sets the curve's shape: small df gives a steep, heavily skewed spike near zero; larger df flattens the peak rightward and pulls the curve toward symmetry. Reading any chi-square value requires its df - the same statistic means different things on different curves.
Mean and spread
The distribution's centre and spread follow the df directly: the mean of a chi-square distribution is
and the standard deviation is \(\sigma = \sqrt{2\,\text{df}}\). At df = 5: mean 5, SD \(\sqrt{10} \approx 3.16\). At df = 20: mean 20, SD \(\sqrt{40} \approx 6.3\). The rules give an instant reading habit - a chi-square statistic of 18 at df 20 is close to centre (ordinary), while the same 18 at df 3 towers more than five SDs above a mean of 3 (extraordinary) - which is why the df must travel with the statistic.
Growth toward symmetry
Formally, a chi-square variable with \(k\) degrees of freedom adds up \(k\) standard normal variables, each squared and each independent of the others:
That explains two facts at once: it can never be negative, and its mean is \(k\), because each squared \(z\) averages 1. As df grows, the peak moves right, the skew eases, and the sum of many pieces drifts toward a bell shape, a central-limit effect. The drift is slow: a common rule of thumb treats the curve as roughly normal, with mean df and SD \(\sqrt{2\,\text{df}}\), only once df is above about 90. At the small df of most classroom tests the skew matters, so tail areas come from a table or software.
Worked example: reading a statistic against its df
Two researchers report chi-square statistics of 18. The first worked a goodness-of-fit test with 3 categories: df = 2, mean 2, SD \(\sqrt{4} = 2\), so 18 sits \((18-2)/2 = 8\) SDs above centre - an enormous departure. The second worked a 20-category test: df = 19, mean 19, SD \(\sqrt{38} \approx 6.2\), so 18 is \((18-19)/6.2 \approx -0.16\) SDs - dead centre, utterly ordinary. Check: identical statistics, opposite verdicts - the df converts a raw number into evidence, and no chi-square value means anything without it.
Worked example: a first statistic
A snack mix claims equal parts of three ingredients. A handful gives observed counts 15, 20, 25 against expected 20 each. 1. Departures: \(-5, 0, +5\). 2. \(\chi^2 = \frac{25}{20} + \frac{0}{20} + \frac{25}{20} = 1.25 + 0 + 1.25 = 2.5\), with df = \(3 - 1 = 2\). 3. Reading: the mean of the \(\text{df} = 2\) curve is 2, and 2.5 sits barely above it - a small departure. 4. Software gives \(p \approx 0.29\): counts at least this uneven occur in about 29% of fair handfuls - no evidence against the claim. Check: the statistic is small, the p-value large, and the verdict "consistent with expectation" follows both.
Common mistakes to avoid
First, quoting a chi-square value without its degrees of freedom - the same number is ordinary on one curve and damning on another. Second, expecting symmetry: small values are near-certain, and the tail runs right only. Third, reaching for a normal approximation at small df, where the skew is steepest. Fourth, negative chi-square values in a calculation: a sign error, since every term is a square. Fifth, mixing the df formulas - each test has its own, and the coming topics derive them. Sixth, interpreting a large p-value as proof the model is right: consistency is not proof.
Key takeaways
The chi-square distribution prices departures of counts from expectations, via \(\chi^2 = \sum (O-E)^2/E\) - squared, scaled gaps, always non-negative, summed over categories. The family has one parameter, the degrees of freedom, which each test supplies: mean \(\text{df}\), SD \(\sqrt{2\,\text{df}}\), right-skewed with the skew easing as df grows. A statistic is read only against its df: 18 is eight SDs out at df 2 and dead centre at df 19. The handful's \(\chi^2 = 2.5\) at df 2 gives \(p \approx 0.29\) - ordinary unevenness. Three tests and a variance test now put the family to work.
Flashcards
Flashcard
The chi-square statistic
Test your understanding
Question 1 of 5
Values of a chi-square statistic are:
Question 2 of 5
A chi-square distribution has df = 5. Its mean and standard deviation are:
Question 3 of 5
The chi-square distribution's shape is:
Question 4 of 5
Observed counts 15, 20, 25 against expected 20 each give a chi-square statistic of:
Question 5 of 5
The statistic of 2.5 at df = 2 gives p ≈ 0.29. At α = 0.05, the verdict is:
