Skip to content
CALX
Appearance

Appearance

Mode


Palette

Chi-square test, goodness of fit and independence

χ² = Σ (O − E)² / E

Two tests share the chi-square name and this tool keeps them apart: goodness of fit, which confronts observed counts with a stated distribution, and independence, which crosses two characters in a table and derives the expected counts from its margins. Each run returns the statistic, its degrees of freedom, the p-value, the critical value at the chosen level, the contribution of every cell and its standardised residual, Cramér’s V for the strength of the relationship, and the check of Cochran’s rule that decides whether the approximation is admissible at all.

Is a die fair, does a colour appear as predicted, does a genetic cross follow 9:3:3:1? You supply the observed counts and the expected distribution.

One number per category, separated by spaces. These are counts, not percentages.

Counts, frequencies or proportions: they are rescaled to the observed total. Leave empty for a uniform distribution.

0 if the expected distribution comes from theory. 1 if you estimated a mean (Poisson), 2 for a mean and a standard deviation (normal): each one costs a degree of freedom.
0.05 by convention; 0.01 for a stricter requirement.

Scientific dossier


What the tool computes, what it assumes, where it stops being valid, and where its data comes from.

Method & formulasχ² = Σ (O − E)² / E

χ² = Σ (O − E)² / E

E(independence) = (row total × column total) / total

df(goodness of fit) = k − 1 − m

df(independence) = (r − 1)(c − 1)

V = √(χ² / (N × min(r − 1, c − 1)))

The statistic adds up, cell by cell, the squared departure between observed and expected count, divided by the expected count. That division is what makes departures comparable across cells: ten too many against an expectation of a thousand does not weigh like ten too many against an expectation of ten. Under the null hypothesis the sum approximately follows a chi-square law whose degrees of freedom count what the data have not already been used to fix.

O
· observed count in a cell.
E
· expected count under the null hypothesis: given for goodness of fit, derived from the margins for independence.
k
· number of categories in the goodness-of-fit test.
m
· number of parameters estimated from the data to build the expected distribution.
r, c
· number of rows and columns of the contingency table.
N
· total count across all cells.
Standardised residual
· (O − E) / √E: it says which cell carries the departure, and in which direction. Beyond 2 in absolute value the local departure is strong.
Cramér’s V
· strength of the relationship between two characters, from 0 to 1, independent of total sample size.
p-value
· probability of a departure at least this large if the null hypothesis were true.
Validity domainThe chi-square law is only an approximation to the true law of the statistic, valid when expected counts are large enough: Cochran’s rule asks that all be at least 1, and that at most a fifth be below 5.

The chi-square law is only an approximation to the true law of the statistic, valid when expected counts are large enough: Cochran’s rule asks that all be at least 1, and that at most a fifth be below 5. The tool checks that condition and says so, without quietly correcting the result: when it fails, the p-value is optimistic and the test rejects too readily, an exact Fisher test being preferable on a 2×2. The test also assumes independent observations and counts, never percentages: entering percentages amounts to declaring a sample of 100 individuals and distorts the p-value. The Yates correction is defined for a 2×2 table only; requested elsewhere it is ignored, and the tool says so.

Reading the resultA p-value below the level α leads to rejecting the null hypothesis: the counts do not follow the stated distribution, or the two characters are not independent.

A p-value below the level α leads to rejecting the null hypothesis: the counts do not follow the stated distribution, or the two characters are not independent. A p-value above the level does not prove the null hypothesis; it only records that this data is not enough to contradict it, which may reflect too small a sample. The test never says which cell makes the difference: that is the job of the cell-by-cell detail and the standardised residual. Nor does it say which character causes the other, since an association may come from a third common factor.