A categorical test has many tiny expected counts. Why not trust the asymptotic chi-square p-value automatically?
StatisticsLevel: IntermediateBinary and categorical inference
A categorical test has many tiny expected counts. Why not trust the asymptotic chi-square p-value automatically?
Hints
- The approximation concerns expected counts.
- Do not merge categories just to obtain significance.
Answer
The asymptotic reference distribution may be inaccurate. Use an appropriate exact or Monte Carlo method, or combine categories only with a substantive justification.
Source and licence: SciPy — contingency tables · BSD-3-Clause · answer: Quantoros editorial