Course contentsShow
← Data Science
Lesson 445 of 2,14512. ANOVA and Multiple ComparisonsFree lesson

Effect Size: Eta-Squared and Omega-Squared

Measuring the proportion of variance explained by group membership using effect size measures.

Effect Size: Eta-Squared and Omega-Squared

What you'll learn: How to measure the proportion of variance in your data explained by group differences using eta-squared and omega-squared.

Why Effect Size Matters in ANOVA

After conducting a one-way ANOVA and finding a significant F-statistic, you know groups differ—but how much do those group differences matter? Effect sizes answer this by telling you what proportion of the total variance in your outcome is explained by group membership.

Eta-Squared (η²)

Eta-squared is the most straightforward effect size for ANOVA. It represents the proportion of total variance explained by your groups:

η² = SS_between / SS_total

Think of it like a pie chart of variance: if η² = 0.25, then 25% of all the variability in your data is due to which group participants belong to, while 75% is due to individual differences within groups.

Interpretation guidelines:

  • Small: η² ≈ 0.01 (1%)
  • Medium: η² ≈ 0.06 (6%)
  • Large: η² ≈ 0.14 (14%)

Omega-Squared (ω²)

Eta-squared tends to overestimate the true effect in the population, especially with small samples. Omega-squared provides a less biased, more conservative estimate by adjusting for sample size.

The calculation is slightly more complex but follows from your ANOVA table values. Omega-squared is always smaller than eta-squared and gives you a better estimate of what you'd find in the broader population.

When to use which: Report eta-squared for descriptive purposes with your current sample; use omega-squared when making inferences about population-level effects.

Key Takeaway: Eta-squared and omega-squared tell you the proportion of variance explained by group membership—turning your significant ANOVA result into a meaningful measure of practical importance. Omega-squared is the preferred, less-biased estimate for population inference.