Identify the estimator
In a one-factor design, η² and partial η² can coincide. In factorial or repeated-measures designs they often differ, with partial η² numerically larger because other effects are excluded from its denominator. Compare values only when design and estimator match.
Suppose an effect has SS = 20, its associated error SS = 80, and total SS = 200. Then η² = 20/200 = .10, while partial η² = 20/(20 + 80) = .20. Both values are correct for their definitions. Calling them the same “percent explained” would hide the denominator difference, and comparing the partial value with an η² value from another paper would exaggerate the contrast.
Published conventions often repeat .01, .06, and .14 as small, medium, and large reference points linked to Cohen-style effect-size guidance.2
These are fallback labels, not universal importance boundaries. Cohen’s effect-size conventions explains the same problem, ANOVA output shows the source sums of squares, and provenance guidance keeps the attribution visible.
Omega squared is often reported as a less biased population-effect estimator. Name the estimator and interval rather than importing η² labels into every effect-size column.
In repeated-measures output, generalized eta squared may be provided to improve comparability across designs. It introduces another denominator rather than repairing all comparisons automatically. Report the exact symbol or software label, calculation convention, design, and uncertainty.
A defensible reporting sentence
“The interaction had partial η² = .08 (95% CI [.02, .15]), a medium value under the cited .01/.06/.14 convention. Practical interpretation used the outcome scale and design rather than the label alone.”