Statistics Glossary → Epsilon squared (ε²)
Epsilon squared (ε²)concept
An effect size that estimates the proportion of variance explained by a factor while correcting the upward bias of η².
Eta squared (η² = SS_between / SS_total) gives the variance explained in the sample but systematically overestimates the population value, because chance differences between groups are also counted as 'explained'. Epsilon squared corrects this excess: ε² = (SS_between - df_between · MS_error) / SS_total. Omega squared (ω²) is a similar correction with an extra MS_error in the denominator; the two usually give very similar values, slightly below η².
In small samples the corrected measures can be noticeably smaller than η²; when the effect is very weak they can even fall below zero, which is usually reported as zero. A rank-based epsilon squared is also defined for the Kruskal-Wallis test: ε² = H / (N - 1); it does not share the ANOVA formula but is interpreted similarly on a scale from 0 to 1.
An everyday example
The effect of three irrigation schedules on tomato yield is tested in 15 greenhouses. η² comes out at 0.30, so 30% of the yield differences in the sample appear to be due to the schedule. Because the sample is small, ε² is about 0.18, a more realistic estimate of the true effect in the population.
In MerQur
The 'Extra output: ε²' checkbox of the Kruskal-Wallis test (ε² = H/(N - 1)); it also appears in the effect-size output of One-Way ANOVA and ANCOVA and in the Effect Size Calculator.
Analyses where this concept appears
One-Way ANOVAANCOVAKruskal-WallisEffect Size Calculator
Related concepts
Eta squared (η²) and partial eta squared (η²ₚ)Omega squared (ω²)