Chi-Square Test

In quality engineering, the chi-square test serves two distinct roles. The goodness-of-fit variant tests whether observed data matches a theoretical distribution by binning continuous measurements into intervals and comparing observed versus expected counts. The independence variant tests whether defect types are associated with specific machines, shifts, or operators.

The independence test is a workhorse for root cause analysis. When you have attribute data — pass/fail, defect type A/B/C, operator 1/2/3 — the chi-square test of independence determines whether the defect distribution differs significantly across factors. This drives investigation priorities: if defect rate is statistically independent of operator, do not waste time on operator retraining.

The goodness-of-fit variant is less ideal for continuous quality data. Binning destroys information about the exact values, and results depend on bin width and placement. For continuous measurements, the Kolmogorov-Smirnov or Anderson-Darling tests are preferred. The chi-square goodness-of-fit test is best reserved for naturally categorical data or large samples where binning loss is minimal.

EntropyStat's approach to distribution fitting makes the chi-square goodness-of-fit test unnecessary for continuous process data. The EGDF constructs a distribution directly from data without binning, preserving all measurement information. Where the chi-square test requires arbitrary decisions about bin count and placement, the EGDF uses entropy-based optimization to find the best distribution representation automatically.

For the independence testing application, EntropyStat's cluster detection via the ELDF provides a complementary perspective. Instead of asking "is defect rate independent of machine?" in a binary yes/no framework, the ELDF can reveal whether measurements from different machines form distinct clusters in the distribution. This answers a richer question: not just whether machines differ statistically, but how their distributions differ in shape, location, and spread.

EntropyStat also handles the small-sample limitation that plagues chi-square tests. The chi-square test requires expected cell counts ≥ 5, which often fails with small production runs or rare defect categories. The EGDF has no minimum count requirement per category — it works with the full continuous dataset regardless of size.