Skip to main content
v2026.11,986 entries · CC-BY 4.0

Eta-Squared Effect Size: Formula, Calculation, and Interpretation

Calculate and interpret eta-squared effect size in ANOVA with exact formulas, Cohen benchmarks, a worked example, and reproducible Excel, R, and Python code.

Written and maintained by CASRAI Editorial Board

Published

On this page

Eta-squared is a sample-based effect size statistic in analysis of variance (ANOVA) that measures the proportion of total variance in a continuous dependent variable accounted for by an experimental factor or grouping predictor. Commonly denoted by the Greek symbol η², it serves as the direct ANOVA analogue to the coefficient of determination (R²) in linear regression, scaling strictly between 0 and 1 (or 0% to 100% of explained variance).

Definition of eta-squared effect size

In quantitative empirical research, statistical significance testing through an analysis of variance (ANOVA) indicates whether group mean differences are unlikely to have arisen solely through random sampling error. However, a significant F-test p-value conveys nothing about the magnitude or practical importance of the observed treatment effect. To evaluate real-world impact, researchers report a standardized effect size metric that quantifies the strength of association between the independent grouping variable and the dependent outcome.

Eta-squared (η²) was historically derived from Karl Pearson’s correlation ratio (η) to evaluate non-linear associations in grouped data. In the general linear model framework, eta-squared represents the ratio of the between-groups treatment variance to the overall variance across all observations. Because it divides the sum of squares for the effect by the total sum of squares, η² reflects the total descriptive variance explained by that specific factor within the sample at hand.

Unlike omnibus significance tests that scale with sample size, eta-squared provides an intuitive metric: an η² of 0.20 indicates that 20% of the total variability in participant scores is attributable to group membership, while the remaining 80% reflects individual differences, measurement error, or unmodeled factors.

Formula for eta-squared with symbol definitions

The standard definitional formula for classical eta-squared is expressed as the ratio of sums of squares:

η² = SS_effect / SS_total

Where the mathematical components are defined as follows:

  • η² (Eta-squared): The sample proportion of total variance explained by the factor under evaluation, bounded between 0.0 and 1.0.
  • SS_effect (Sum of Squares for the Effect): Also termed SS_between or SS_treatment, this quantifies the dispersion of group sample means around the grand mean across all observations: SS_effect = Σ [ n_i × (M_i – M_grand)² ], where n_i is the sample size of group i, M_i is the mean of group i, and M_grand is the grand mean.
  • SS_total (Total Sum of Squares): The total dispersion of individual observation scores around the grand mean across the entire dataset: SS_total = Σ (X_ij – M_grand)², which identically equals the sum of SS_effect and the residual within-groups error sum of squares (SS_error).
  • SS_error (Sum of Squares Error): Also designated SS_within or SS_residual, representing unexplained variation within individual treatment conditions: SS_error = Σ [ (n_i – 1) × s_i² ], where s_i² is the sample variance of group i.

When raw sums of squares are not directly printed in published literature, researchers can compute eta-squared directly from standard ANOVA test statistics—specifically the omnibus F-ratio and its corresponding degrees of freedom. In a one-way between-subjects design, the algebraic identity is:

η² = (F × df_effect) / [ (F × df_effect) + df_error ]

Where:

  • F: The observed F-statistic from the ANOVA summary table, defined as Mean Square Effect divided by Mean Square Error (MS_effect / MS_error).
  • df_effect: The numerator degrees of freedom for the factor, calculated as k – 1, where k is the number of experimental groups.
  • df_error: The denominator degrees of freedom for the residual error term, calculated as N – k, where N represents total sample size across all conditions.

Worked example: calculating eta-squared step-by-step

To demonstrate the exact arithmetic, consider an experimental study evaluating three instructional interventions on university student statistical comprehension scores (measured on a standardized 0 to 100 scale). A cohort of N = 45 undergraduate researchers is randomly assigned in equal allocations (n = 15 per condition) to one of three instructional formats: Group 1 receives traditional lectures, Group 2 participates in interactive data workshops, and Group 3 utilizes self-paced simulation software.

The descriptive performance metrics across the experimental conditions are recorded as follows:

  • Group 1 (Traditional lecture): n₁ = 15, Mean (M₁) = 72.00, Sum of Squares (SS₁) = 540.00
  • Group 2 (Interactive workshop): n₂ = 15, Mean (M₂) = 78.00, Sum of Squares (SS₂) = 570.00
  • Group 3 (Simulation software): n₃ = 15, Mean (M₃) = 84.00, Sum of Squares (SS₃) = 630.00

Step 1: Calculate the Grand Mean (M_grand)
Because the group sample sizes are balanced (n₁ = n₂ = n₃ = 15):
M_grand = (72.00 + 78.00 + 84.00) / 3 = 234.00 / 3 = 78.00

Step 2: Calculate Between-Groups Sum of Squares (SS_effect)
SS_effect = Σ [ n_i × (M_i – M_grand)² ]
SS_effect = 15 × (72.00 – 78.00)² + 15 × (78.00 – 78.00)² + 15 × (84.00 – 78.00)²
SS_effect = 15 × (-6.00)² + 15 × (0.00)² + 15 × (6.00)²
SS_effect = 15 × 36.00 + 0.00 + 15 × 36.00 = 540.00 + 540.00 = 1,080.00

Step 3: Calculate Within-Groups Error Sum of Squares (SS_error)
SS_error = SS₁ + SS₂ + SS₃ = 540.00 + 570.00 + 630.00 = 1,740.00

Step 4: Calculate Total Sum of Squares (SS_total)
SS_total = SS_effect + SS_error = 1,080.00 + 1,740.00 = 2,820.00

Step 5: Compute Degrees of Freedom and Mean Squares
df_effect = k – 1 = 3 – 1 = 2
df_error = N – k = 45 – 3 = 42
MS_effect = SS_effect / df_effect = 1,080.00 / 2 = 540.00
MS_error = SS_error / df_error = 1,740.00 / 42 = 41.43

Step 6: Compute the Omnibus F-Statistic and Significance
F = MS_effect / MS_error = 540.00 / 41.4286 = 13.03
At df₁ = 2 and df₂ = 42, the critical F threshold at α = .05 is 3.22. The observed F(2, 42) = 13.03 yields an exact p-value of 3.95 × 10⁻⁵ (p < .001), indicating statistically significant performance divergence among instructional methods.

Step 7: Calculate Eta-Squared (η²)
η² = SS_effect / SS_total = 1,080.00 / 2,820.00 = 0.3830 (38.30%)
Verifying via the F-statistic conversion formula yields the identical result:
η² = (13.0345 × 2) / [ (13.0345 × 2) + 42 ] = 26.0690 / (26.0690 + 42) = 26.0690 / 68.0690 = 0.3830

Step 8: Construct Noncentral F 90% Confidence Intervals
Using the noncentral F inversion method required by statistical reporting standards, the noncentrality parameter (λ) bounds for F = 13.03 with df₁ = 2 and df₂ = 42 are λ_lower = 9.3904 and λ_upper = 47.3295. Converting λ to variance proportion yields a 90% confidence interval of [.17, .51]. Thus, the researcher can conclude with 90% confidence that instructional format accounts for between 17% and 51% of variance in statistical comprehension.

Eta squared effect size ranges and interpretation

Interpreting an eta squared effect size requires contextual grounding. In his foundational 1988 treatise on statistical power analysis, Jacob Cohen proposed general operational heuristics to assist behavioral scientists when field-specific baseline distributions are unavailable. Cohen did not define benchmarks arbitrarily; rather, he derived them algebraically from his standardized mean difference metric for multi-group designs, Cohen’s f, using the exact transformation η² = f² / (1 + f²).

By establishing operational benchmarks for Cohen’s f of 0.10 (small), 0.25 (medium), and 0.40 (large), Cohen produced the widely cited eta squared effect size ranges:

Magnitude Cohen’s f Eta-Squared (η²) Percent of Variance Practical Meaning in Research
Small 0.10 ≈ 0.01 1.0% Noticeable across large populations; subtle individual difference; typical in epidemiological screens.
Medium 0.25 ≈ 0.06 5.9% (≈ 6.0%) Visually detectable by careful observation; standard experimental manipulation in psychology and education.
Large 0.40 ≈ 0.14 13.8% (≈ 14.0%) Substantial practical divergence; gross observable difference between conditions; decisive institutional impact.

In our worked example above, the obtained η² of .38 exceeds the 0.14 threshold, indicating an exceptionally large treatment effect wherein 38.3% of the variability in test performance is directly explained by the type of instruction.

Crucially, Jacob Cohen explicitly warned against using these heuristic thresholds as rigid dogmatic rules. The substantive meaning of an effect size depends intrinsically on scientific context, study design, and practical utility. In genome-wide association studies (GWAS) or public health population interventions, an η² of 0.005 (0.5% variance explained) can have profound clinical or economic ramifications across millions of individuals. Conversely, in highly controlled laboratory bench experiments with tight physical constraints, an η² below 0.10 might represent unacceptably high experimental noise.

Classical eta-squared vs. partial and generalized eta-squared

A frequent source of confusion in quantitative empirical literature is the distinction between classical eta-squared (η²), partial eta-squared (η_p²), and generalized eta-squared (η_G²). In a simple one-way between-subjects ANOVA, classical eta-squared and partial eta-squared are mathematically identical because there is only one factor and one error term: SS_total = SS_effect + SS_error.

However, as explored in detailed reviews of effect sizes in analysis of variance, in multi-factor factorial designs (such as a two-way ANOVA with factors A, B, and an A×B interaction), the two metrics diverge fundamentally:

  • Classical Eta-Squared: η² = SS_effect / SS_total. The denominator remains the total variance of the entire experiment. Consequently, classical eta-squared values across all factors and interactions are strictly additive: η²_A + η²_B + η²_AB = R²_model ≤ 1.0.
  • Partial Eta-Squared: η_p² = SS_effect / (SS_effect + SS_error). The denominator excludes the variance accounted for by other systematic factors, isolating the effect against residual error alone. Because each factor’s calculation ignores the variance absorbed by other factors, the sum of partial eta-squared values across effects frequently exceeds 1.0 (100%), precluding additive interpretation. For a direct head-to-head structural analysis, see the dedicated guide on eta-squared versus partial eta-squared.

As documented by methodological historians Levine and Hullett (2002) and Richardson (2011), widespread reporting errors arose because early editions of major software packages (such as SPSS Version 9) erroneously labeled partial eta-squared output simply as “Eta Squared”. As a consequence, thousands of published studies claimed to report classical η² while actually presenting inflated η_p² values, distorting subsequent meta-analyses.

To establish cross-study comparability across diverse experimental designs (such as between-subjects vs. repeated-measures designs), Olejnik and Algina (2003) and Bakeman (2005) introduced generalized eta-squared (η_G²). Generalized eta-squared distinguishes between manipulated factors and individual difference variables, providing a standardized denominator that enables valid meta-analytic synthesis across disparate laboratory protocols. The data table below compares variance-explained effect size statistics across ANOVA designs, detailing mathematical denominators, additivity, sample estimator bias, Cohen benchmark thresholds, and cross-design comparability.

Assumptions, sample size bias, and common mistakes

Valid calculation and interpretation of eta-squared rely on the underlying assumptions of the general linear model:

  1. Independence of Observations: Data points must be mutually independent; clustered or repeated measures require multi-level or repeated-measures ANOVA formulations.
  2. Normality of Residuals: Model residuals within each factor level should follow an approximately normal distribution.
  3. Homogeneity of Variance (Homoscedasticity): Population variances across experimental groups must be approximately equal. When heteroscedasticity is severe, standard sum of squares partitioning is compromised.

Beyond distribution assumptions, quantitative researchers must avoid four widespread methodological pitfalls:

  • Sample Size Upward Bias: Classical eta-squared is a purely descriptive sample statistic, not an unbiased estimator of population variance explained. In small samples, sample means inevitably deviate from one another due to sampling error alone, causing sample SS_effect to be positive even when the true population effect is zero. To estimate population variance explained without sample-size inflation, researchers calculate omega-squared (ω²) or epsilon-squared (ε²), which incorporate degrees of freedom and error mean squares: ω² = [ SS_effect – (df_effect × MS_error) ] / (SS_total + MS_error). In our worked example, while sample η² was .38, population omega-squared adjusts downward to ω² = .35.
  • Equating Explained Variance with Causality: A substantial eta-squared demonstrates strong statistical association within the study cohort but does not establish causal mechanisms in non-experimental or quasi-experimental settings.
  • Mislabeled Software Outputs: Blindly copying effect size tables from statistical software without confirming whether the program defaults to classical, partial, or generalized eta-squared.
  • Neglecting Confidence Intervals: Reporting point estimates of η² without corresponding confidence intervals, which obscures the statistical precision of the finding.

How to report eta-squared in APA style

The Publication Manual of the American Psychological Association (7th edition, 2020, Sections 6.33–6.35) outlines explicit standards for reporting ANOVA outcomes and effect size statistics in academic manuscripts:

  • Symbol Formatting: Greek letters (such as η and η²) must remain unitalicized. Roman test statistics (such as F, p, and N) must be italicized.
  • Leading Zero Rule: Because eta-squared is mathematically bounded between 0 and 1 and cannot exceed 1.0, omit the leading zero before the decimal point (e.g., write η² = .38, never η² = 0.38).
  • Decimal Precision: Report effect size estimates and confidence interval limits to two decimal places.
  • Confidence Interval Convention: Because the F-distribution is strictly non-negative (one-sided test), report a 90% confidence interval for η² rather than a 95% interval. A 90% confidence interval around an F-based effect size corresponds directly to a two-sided α = .05 significance test, avoiding anomalies where a significant test at p < .05 yields a confidence interval overlapping zero.

Standard APA 7th Narrative Reporting Example:
“A one-way between-subjects analysis of variance was conducted to assess the effect of instructional intervention format on statistical comprehension scores. The omnibus test revealed a statistically significant difference between instructional conditions, F(2, 42) = 13.03, p < .001, η² = .38, 90% CI [.17, .51]. Post hoc pairwise comparisons with Tukey HSD adjustment indicated that students utilizing self-paced simulation software (M = 84.00, SD = 6.71) scored significantly higher than those attending interactive workshops (M = 78.00, SD = 6.38, p = .028) and traditional lectures (M = 72.00, SD = 6.21, p < .001)."

How to calculate eta-squared in Excel, R, and Python

Researchers can implement reproducible calculations of eta-squared across standard computational environments using either raw sums of squares or model output summaries.

Calculating eta-squared in Microsoft Excel

When using Excel’s native ANOVA data analysis toolpack or custom spreadsheet formulas, calculate eta-squared directly from the ANOVA output table:

  1. Run Data > Data Analysis > Anova: Single Factor on your grouped columns.
  2. Identify the Between Groups Sum of Squares cell (for example cell B2) and the Total Sum of Squares cell (for example cell B4).
  3. Enter the direct quotient formula in an adjacent cell: =B2/B4.
  4. Alternatively, if only the F-statistic and degrees of freedom are available (with F in E2, df_between in C2, and df_within in C3), calculate η² with: =(E2*C2)/((E2*C2)+C3).

Calculating eta-squared in R

In the R statistical computing environment, compute classical eta-squared using base functions or the specialized effectsize library:

# Base R calculation from an aov object
data <- data.frame(
  method = factor(rep(c("Lecture", "Workshop", "Simulation"), each = 15)),
  score = c(68, 72, 75, 71, 74, 70, 73, 72, 71, 76, 70, 69, 74, 73, 72,
            76, 78, 80, 77, 79, 75, 78, 81, 77, 78, 80, 76, 79, 77, 79,
            83, 85, 82, 86, 84, 85, 87, 83, 84, 86, 82, 85, 84, 83, 85)
)

model <- aov(score ~ method, data = data)
anova_summary <- summary(model)[[1]]

# Extract sums of squares and compute eta-squared
ss_effect <- anova_summary["method", "Sum Sq"]
ss_total <- sum(anova_summary[, "Sum Sq"])
eta_sq <- ss_effect / ss_total
cat(sprintf("Base R Eta-Squared: %.4fn", eta_sq))

# Using the effectsize package for point estimates and 90% CIs
if (requireNamespace("effectsize", quietly = TRUE)) {
  library(effectsize)
  res <- eta_squared(model, partial = FALSE, ci = 0.90)
  print(res)
}

Calculating eta-squared in Python

In Python, calculate eta-squared utilizing pandas and statsmodels for ordinary least squares modeling:

import pandas as pd
import statsmodels.api as sm
from statsmodels.formula.api import ols

# Construct sample dataframe
data = {
    'method': ['Lecture']*15 + ['Workshop']*15 + ['Simulation']*15,
    'score': [72, 71, 73, 70, 74, 71, 72, 73, 72, 71, 73, 72, 71, 74, 71,
              77, 78, 79, 76, 80, 77, 78, 79, 78, 77, 79, 78, 77, 80, 77,
              84, 83, 85, 82, 86, 83, 84, 85, 84, 83, 85, 84, 83, 86, 83]
}
df = pd.DataFrame(data)

# Fit ANOVA model
model = ols('score ~ C(method)', data=df).fit()
anova_table = sm.stats.anova_lm(model, typ=2)

# Compute classical eta-squared
ss_effect = anova_table.loc['C(method)', 'sum_sq']
ss_total = anova_table['sum_sq'].sum()
eta_squared = ss_effect / ss_total

print(f"ANOVA Summary:n{anova_table}")
print(f"Calculated Eta-Squared (η²): {eta_squared:.4f}")

Frequently asked questions

What is the difference between eta-squared and partial eta-squared?

Classical eta-squared (η²) divides the effect sum of squares by the total experiment sum of squares (SS_effect / SS_total), representing the absolute proportion of total variance explained. In contrast, partial eta-squared (η_p²) divides the effect sum of squares by the sum of that effect and residual error alone [SS_effect / (SS_effect + SS_error)], ignoring variance absorbed by other factors in the model. In a one-way ANOVA with a single between-subjects factor, η² and η_p² yield identical numbers; in factorial designs, partial eta-squared produces larger values that are non-additive.

What are the benchmark ranges for interpreting eta-squared effect size?

Under Jacob Cohen’s widely used 1988 operational conventions, an eta-squared value of approximately .01 (1% variance explained) is interpreted as small, .06 (6% variance explained) is considered medium, and values of .14 or greater (14% or more variance explained) are classified as large. However, researchers should treat these values as guidelines of last resort rather than rigid thresholds, interpreting findings relative to empirical literature in their specific disciplinary field.

Why is eta-squared considered a biased estimator of population effect size?

Eta-squared is a descriptive sample statistic that always overestimates the true proportion of variance explained in the population, particularly in small samples. Because sample means will inevitably vary by chance alone, the sample sum of squares for an effect is always positive. For an unbiased estimate of population variance explained that penalizes for model complexity, researchers report omega-squared (ω²) or epsilon-squared (ε²).

Can eta-squared exceed 1.0 or 100 percent in an ANOVA?

No. Classical eta-squared is mathematically bounded between 0.0 and 1.0 (0% to 100%) because the effect sum of squares cannot exceed the total sum of squares. Furthermore, in factorial ANOVA, the sum of classical eta-squared values across all main effects and interaction terms cannot exceed 1.0. By contrast, partial eta-squared values are non-additive and their sum across multiple factors can exceed 100%.

Why does APA style require a 90 percent confidence interval for eta-squared instead of 95 percent?

APA 7th edition guidelines recommend reporting a 90% confidence interval for ANOVA effect sizes because the omnibus F-test is fundamentally a non-negative, one-tailed test. In an F-distribution, all variance deviations produce positive test statistics. Constructing a 90% confidence interval corresponds directly to a two-sided α = .05 significance test, ensuring that whenever an omnibus F-test is statistically significant at p < .05, the lower bound of the confidence interval will be strictly greater than zero.

Sources

  • American Psychological Association. (2020). Publication Manual of the American Psychological Association (7th ed.). https://apastyle.apa.org/style-grammar-guidelines/statistics/ (Accessed 2026-11-15).
  • Bakeman, R. (2005). Recommended effect size statistics for repeated measures designs. Behavior Research Methods, 37(3), 379–384. https://doi.org/10.3758/BF03192707 (Accessed 2026-11-15).
  • Cohen, J. (1988). Statistical Power Analysis for the Behavioral Sciences (2nd ed.). Lawrence Erlbaum Associates. https://www.utstat.toronto.edu/~brunner/oldclass/378f16/readings/CohenPower.pdf (Accessed 2026-11-15).
  • Lakens, D. (2013). Calculating and reporting effect sizes to facilitate cumulative science: A practical primer for t-tests and ANOVAs. Frontiers in Psychology, 4, 863. https://doi.org/10.3389/fpsyg.2013.00863 (Accessed 2026-11-15).
  • Levine, T. R., and Hullett, C. R. (2002). Eta squared, partial eta squared, and misreporting of effect size in communication research. Human Communication Research, 28(4), 612–625. https://doi.org/10.1111/j.1468-2958.2002.tb00828.x (Accessed 2026-11-15).
  • Olejnik, S., and Algina, J. (2003). Generalized eta and omega squared statistics: Measures of effect size for some common research designs. Psychological Methods, 8(4), 434–447. https://doi.org/10.1037/1082-989X.8.4.434 (Accessed 2026-11-15).
  • Richardson, J. T. E. (2011). Eta squared and partial eta squared as measures of effect size in educational research. Educational Research Review, 6(2), 135–147. https://doi.org/10.1016/j.edurev.2010.12.001 (Accessed 2026-11-15).

Data: Comparison of ANOVA Variance-Explained Effect Size Statistics

Comparative compilation of variance-explained effect size statistics in ANOVA models, detailing mathematical denominators, additivity across model factors, sample estimator bias properties, Cohen benchmark thresholds, and cross-design comparability compiled from Cohen (1988), Olejnik and Algina (2003), Richardson (2011), and Lakens (2013).

Comparison of ANOVA Variance-Explained Effect Size Statistics
StatisticDenominator FormulationAdditive PropertySample Estimator BiasCohen Small Cutoff (proportion)Cohen Medium Cutoff (proportion)Cohen Large Cutoff (proportion)Design Comparability
Eta-Squared (η²)SS_totalYes (sums to R² ≤ 1.0)Upwardly biased (overestimates population effect)0.010.060.14Between-subjects designs only
Partial Eta-Squared (ηp²)SS_effect + SS_errorNo (sum of effects can exceed 1.0)Upwardly biased (overestimates population effect)0.010.060.14Limited (design-dependent error structures)
Generalized Eta-Squared (ηG²)SS_effect + SS_measured + SS_errorDesign dependentUpwardly biased (overestimates population effect)0.010.060.14High (designed specifically for cross-design meta-analyses)
Omega-Squared (ω²)SS_total + MS_errorYes (sums to estimated population R²)Less biased (adjusted for degrees of freedom and error variance)0.010.060.14Between-subjects designs only
Partial Omega-Squared (ωp²)SS_effect + (N - df_effect) * MS_errorNo (non-additive across model factors)Less biased (adjusted for degrees of freedom and error variance)0.010.060.14Limited (cannot generalize across mixed designs)
Epsilon-Squared (ε²)SS_totalYes (sums to adjusted sample variance)Less biased (less sample bias than classical eta-squared)0.010.060.14Between-subjects designs only

Sources

Compiled by CASRAI from the sources above. Free to reuse under CC BY 4.0 with attribution.

Follow CASRAI

Research-administration guidance, standards updates and independent tool reviews.

Ask CASRAI · free to try

Ask about Eta-Squared Effect Size: Formula, Calculation, and Interpretation

2 free questions a day, no account.

An AI assistant for research: funder rules, compliance, methods and publishing. It links the sources behind its answers and labels answers that come from a web search.

Answers draw on CASRAI's guides and dictionary plus the federal and funder documents we index: Federal Register, Grants.gov, Regulations.gov and UKRI.

Works on this site and inside Claude, Cursor and the AI tools you already use.

Everything CASRAI publishes — this page, the dictionary, the guides and the news — stays open to read, with no account and no card.

Referenced across the research world

University of Cambridge logoColumbia University logoCrossref logoUniversity of Edinburgh logoHarvard University logoUniversity of Oxford logoPrinceton University logoStanford School of Medicine logoUniversity College London logoORCID logoUniversity of Cambridge logoColumbia University logoCrossref logoUniversity of Edinburgh logoHarvard University logoUniversity of Oxford logoPrinceton University logoStanford School of Medicine logoUniversity College London logoORCID logo
  • University of Cambridge logo
  • Columbia University logo
  • Crossref logo
  • University of Edinburgh logo
  • Harvard University logo
  • University of Oxford logo
  • Princeton University logo
  • Stanford School of Medicine logo
  • University College London logo
  • ORCID logo

View CASRAI adoption →