1 Definition and interpretation
Eta squared is an effect size statistic used in analysis of variance and related procedures. It expresses the share of variability in a dependent variable that is associated with a given factor, treatment, or grouping variable. Unlike a significance test, which asks whether an effect is likely to be different from zero, eta squared describes the magnitude of that effect.
1.1 Basic concept
In an ANOVA setting, observations vary because of differences between groups and because of random variation within groups. Eta squared summarizes how much of the total variation can be attributed to the factor being studied. A larger value suggests that group membership or treatment condition is more strongly associated with the outcome.
1.2 Proportion of variance explained
Eta squared is interpreted as a proportion. For example, an eta squared of 0.20 indicates that 20% of the total variance in the dependent variable is linked to the factor in the model. The remaining variance is due to other factors, measurement error, or unexplained differences among observations.
1.3 Range and meaning of values
The statistic ranges from 0 to 1 in standard one-way settings. A value near 0 indicates little association between the factor and the outcome, while a value near 1 indicates that most of the variance is accounted for by the factor. In practice, values must be interpreted in context, since even a small proportion can be meaningful in some fields.
1.4 Relationship to effect size
Eta squared belongs to the broader family of effect size measures. It is used to supplement p-values by showing the substantive strength of an observed effect. Because it focuses on magnitude rather than statistical significance, it is especially useful for comparing findings across studies or conditions.
2 Calculation
Eta squared is derived from sums of squares in an ANOVA table. It compares the variability explained by the effect of interest with the total observed variability.
2.1 Formula for eta squared
The standard formula is:
η² = SSbetween / SStotal
where SSbetween is the sum of squares associated with the factor or treatment, and SStotal is the total sum of squares.
2.2 Sum of squares components
The numerator reflects variability explained by the factor, while the denominator includes both explained and unexplained variability. In a simple one-way ANOVA, SSbetween captures differences among group means, and SSwithin captures variation among observations inside groups. Since SStotal = SSbetween + SSwithin, eta squared can be read as the proportion of total variation due to group differences.
2.3 Computation from ANOVA tables
ANOVA tables usually list sums of squares for each source of variation. To compute eta squared, divide the sum of squares for the effect by the total sum of squares. In designs with more than one factor, the choice of which effect to use in the numerator depends on the research question and the model structure.
2.4 Worked example
Suppose an ANOVA table reports SSbetween = 24 and SStotal = 80. Then:
η² = 24 / 80 = 0.30
This result indicates that 30% of the total variance in the outcome is associated with the factor under study. The remaining 70% is attributed to within-group variation and other unexplained influences.
3 Variants and related measures
Several closely related statistics are used in place of or alongside eta squared, especially in more complex designs.
3.1 Partial eta squared
Partial eta squared estimates the proportion of variance explained by a particular effect relative to that effect plus its associated error term. It is widely reported in factorial and repeated-measures designs. Because its denominator excludes some sources of variance included in eta squared, it often yields larger values.
3.2 Generalized eta squared
Generalized eta squared is designed for comparing effects across different experimental designs. It uses a denominator that incorporates additional variance components, making it more comparable across studies with different structures. It is often recommended when one wants a measure that is less dependent on the specific design.
3.3 Omega squared
Omega squared is another effect size estimate for ANOVA. It adjusts for sample bias and tends to provide a more conservative estimate of population effect size than eta squared. For this reason, some statisticians prefer it when the goal is to estimate explained variance more realistically.
3.4 Comparison with other effect size measures
Eta squared is most directly tied to variance partitioning in ANOVA. Other effect size measures, such as Cohen’s d or r, are used in different settings or provide different summaries of association. Conversions between measures are sometimes possible, but the choice of statistic should match the design and the research question.
4 Use in statistical analysis
Eta squared is commonly reported in experimental and quasi-experimental studies as a descriptive index of effect magnitude.
4.1 One-way ANOVA
In a one-way ANOVA, eta squared is straightforward to compute and interpret. It summarizes how much of the total outcome variance is associated with differences among the group means. This makes it a common companion to the F test.
4.2 Factorial ANOVA
In factorial designs, each main effect and interaction can be assigned its own effect size estimate. Eta squared or partial eta squared may be reported for each term to show the relative importance of factors and their combined influence on the dependent variable.
4.3 Repeated-measures designs
Repeated-measures analyses often use partial eta squared or generalized eta squared rather than the simplest form of eta squared. This is because the same participants contribute multiple observations, and the dependence structure affects how variance should be partitioned. The chosen statistic should reflect the design clearly.
4.4 Reporting results
A typical report includes the F statistic, degrees of freedom, p-value, and an effect size such as eta squared. Reporting effect size helps readers judge the practical magnitude of the finding, not just whether it is statistically detectable. Authors should also specify which variant was used, since different forms are not directly interchangeable.
5 Interpretation and guidelines
5.1 Conventional benchmarks
Some fields use rough benchmarks such as small, medium, and large effects. These rules of thumb are only guides and should not be treated as universal standards. A value considered small in one discipline may be consequential in another.
5.2 Practical significance
Statistical effect size does not automatically imply practical importance. An eta squared value may be modest yet still matter in applied contexts where small changes have large consequences. Interpretation should consider the nature of the outcome, the cost of intervention, and the broader research setting.
5.3 Limitations in interpretation
Eta squared does not indicate causality on its own, even when used in experimental analyses. It also depends on the design and the set of variables included in the model. Comparisons across studies can be misleading if the structures, measurement scales, or sources of variance differ substantially.
6 Advantages and limitations
6.1 Strengths of eta squared
Eta squared is easy to compute from standard ANOVA output and has a clear interpretation as explained variance. It is intuitive for readers and can help translate statistical results into a more accessible form. Its simplicity has made it a longstanding descriptive measure in the social and behavioral sciences.
6.2 Common criticisms
A frequent criticism is that eta squared can be upwardly biased, especially in smaller samples. It may also overstate the importance of effects in certain designs because it does not adjust for all sources of estimation error. As a result, it can present a somewhat optimistic picture of effect magnitude.
6.3 Bias and sample size considerations
Because sample size affects estimation precision, small studies may produce unstable values of eta squared. Large samples can yield more stable estimates, but even then, the statistic remains sample-dependent. Researchers often complement it with confidence intervals or alternative measures to improve interpretation.
6.4 When to prefer alternative measures
Alternative statistics may be preferable when a design is complex, when a less biased estimate is needed, or when comparison across studies is important. Omega squared is often selected for bias adjustment, while generalized eta squared can be useful in mixed or repeated-measures designs. The best choice depends on the analytic context and reporting conventions.
7 Applications in research
Eta squared is widely used in empirical disciplines that rely on group comparisons and experimental designs.
7.1 Psychology
Psychology frequently uses eta squared to report the size of effects in experiments, intervention studies, and laboratory tasks. It helps describe the strength of treatment conditions, individual differences, and interaction effects.
7.2 Education
In education research, eta squared may summarize the influence of instructional methods, classroom conditions, or assessment formats on student outcomes. It is useful for indicating whether a teaching approach has a modest or substantial association with achievement measures.
7.3 Social sciences
Sociology, communication studies, and related fields use eta squared to quantify how much variance in attitudes, behaviors, or survey outcomes is linked to categorical factors. The statistic supports clearer interpretation of group differences in observational and experimental work.
7.4 Other quantitative fields
Eta squared also appears in health research, business analytics, and other areas that use ANOVA-type models. In these settings, it serves as a compact summary of effect magnitude and a supplement to hypothesis testing.