What Is a Chi-Square Test? Definition and Uses

The chi-square test is the workhorse for count data — observations sorted into categories rather than measured on a scale. It compares the counts you observed against the counts you would expect if nothing interesting were happening, and asks whether the mismatch is too large to be chance. When your data are defects by type, errors by shift, or returns by reason, this is the test that fits.

How it’s used

Two uses dominate. A goodness-of-fit test checks whether one set of counts matches a claimed distribution — are defects really spread evenly across five categories? A test of independence, the more common in improvement work, checks whether two categorical variables are related: does defect type depend on machine, does error rate depend on shift, does complaint category depend on region? The data are laid out in a contingency table, expected counts are computed for each cell under the assumption of no relationship, and the chi-square statistic totals up the discrepancies. A large total, and its small p-value, says the pattern is real.

The test needs raw counts, not percentages, and enough data that expected counts are not tiny — a common rule of thumb asks for at least five expected observations per cell. Sparse tables call for combining categories or switching to an exact method.

A worked example

Imagine a distribution center logging picking errors across three shifts for a month, with each error classified as wrong item, wrong quantity, or damaged goods. Laid out as a three-by-three table, the counts look uneven — night shift seems heavy on wrong-item errors. The chi-square test of independence asks whether error type and shift are genuinely related, or whether a table this uneven could plausibly arise by chance. A small p-value would tell the team the pattern deserves investigation — perhaps night shift picks from poorly lit aisles, or covers unfamiliar zones. The test finds the association; the walk to the floor finds the cause.

• Feed the test raw counts — running it on percentages or rates produces nonsense.
• Watch expected cell counts; below about five per cell, combine categories or use an exact method.
• Association is not causation — a significant result says related, never why.
• Count each unit once; letting one order contribute errors to multiple cells breaks independence.
• With very large samples, trivial imbalances turn significant — read the table, not just the p-value.

When the data are counts, the chi-square test is how you ask whether the pattern would survive a coin flip.

The chi-square test sits alongside the t-test and ANOVA in the Analyze toolkit of our Green Belt program — $299, 35 hours, and a 100-question proctored exam with a 70 percent pass mark and one free retake. It is often the first formal test a team runs, because defect data usually arrive as counts.