Written and maintained by CASRAI Editorial Board
Published
On this page
Effect size in ANOVA is a number that says how much of the variation in the outcome is attributable to a factor or interaction, not just whether the F test is significant. The most common indices are eta squared (η²), partial eta squared (ηp²) and omega squared (ω²), each a proportion of variance, and Cohen’s f, which is a standardised spread of the group means. This guide gives the formulas with every symbol defined, one fully worked two-factor example, benchmark caveats, a reporting pattern and code in Excel, R and Python.
What effect size means in ANOVA
ANOVA’s F test answers whether the group means differ by more than chance. An effect size answers a different question: how large is the difference relative to the other variation in the data? With a large enough sample a trivial difference is significant, and with a small sample a large difference can fail to be, so a p-value alone cannot tell a reader how much a factor matters. Lakens (2013) states the practical rule plainly: always report effect sizes, and say which version you are reporting.
ANOVA effect sizes fall into two families. Variance-explained indices (η², ηp², ω²) express the share of variability attributable to an effect. Cohen’s f expresses the standard deviation of the group means in units of the within-group standard deviation, which is the form used in power analysis. The two families are linked by simple algebra, shown below. For the general concept outside ANOVA see the effect size dictionary entry.
Formulas for ANOVA effect size
Notation: SSeffect is the sum of squares for the factor or interaction, SSerror the residual (within-cell) sum of squares, SStotal the total sum of squares, dfeffect its degrees of freedom, MSeffect = SSeffect / dfeffect, MSerror = SSerror / dferror, and N the total number of observations. The formulas below are those given in Lakens (2013), equations 11 to 15, with Cohen’s f from the same article.
- Eta squared: η² = SSeffect / SStotal. It is the proportion of total sample variance explained by the effect, and it is uncorrected, so it overstates the population value.
- Partial eta squared: ηp² = SSeffect / (SSeffect + SSerror). For fixed factors this equals F × dfeffect / (F × dfeffect + dferror), so it can be recovered from a published F test.
- Omega squared (between-subjects): ω² = dfeffect × (MSeffect − MSerror) / (SStotal + MSerror).
- Partial omega squared: ωp² = dfeffect × (MSeffect − MSerror) / (dfeffect × MSeffect + (N − dfeffect) × MSerror).
- Cohen’s f: f² = ηp² / (1 − ηp²), so f = √(ηp² / (1 − ηp²)), and conversely ηp² = f² / (1 + f²).
Effect size of ANOVA for one factor versus several
In a one-way ANOVA there is only one effect, so η² and ηp² are identical. In a design with two or more factors they diverge, because ηp² removes the variance explained by the other factors from the denominator while η² does not. Lakens notes that when every factor is manipulated between participants, generalized eta squared and partial eta squared are identical; the generalized version matters when designs include measured covariates or repeated measures. The eta squared versus partial eta squared comparison walks through that difference in more detail.
Worked example: effect size for a two-factor ANOVA
This is an illustrative calculation with invented numbers, not data from a study. Suppose a balanced 2 × 3 between-subjects design, with factor A at two levels and factor B at three levels, 10 observations per cell, so N = 60. The ANOVA table is:
| Source | SS | df | MS | F |
|---|---|---|---|---|
| A | 48.0 | 1 | 48.0 | 6.00 |
| B | 90.0 | 2 | 45.0 | 5.63 |
| A × B | 30.0 | 2 | 15.0 | 1.88 |
| Error | 432.0 | 54 | 8.0 | |
| Total | 600.0 | 59 |
Here dferror = 60 − 6 = 54, MSerror = 432 / 54 = 8.0, and each F is MSeffect / 8.0. The arithmetic for factor B, step by step:
- η² = 90 / 600 = 0.150.
- ηp² = 90 / (90 + 432) = 90 / 522 = 0.172.
- ω² = 2 × (45 − 8) / (600 + 8) = 74 / 608 = 0.122.
- ωp² = 2 × (45 − 8) / (2 × 45 + 58 × 8) = 74 / 554 = 0.134.
- f² = 0.172 / (1 − 0.172) = 0.208, so f = 0.456.
- Check from F: 5.625 × 2 / (5.625 × 2 + 54) = 11.25 / 65.25 = 0.172, matching step 2.
The data table below applies the same formulas to A, B and A × B. Two patterns are worth reading off it. First, ω² is smaller than η² for every effect, as expected for a less biased estimate. Second, A × B has an F of 1.88, which is not significant at the usual level (the p-value is about .16), and it also has the smallest effect sizes, so the effect size and the test tell the same story here.
How to interpret ANOVA effect size
Cohen’s (1988) conventions, as relayed by Lakens (2013) and by the documentation of the R package effectsize, are the reference most readers expect. The f values are quoted by effectsize as Cohen’s suggestion; the η² values are those Lakens lists. The corresponding f for each η² is converted with the formula above.
| Label | η² (Cohen 1988) | Cohen’s f |
|---|---|---|
| Small | .01 | .10 |
| Medium | .06 | .25 |
| Large | .14 | .40 |
In the worked example, factor B has η² = .15 and f = .46, which sits just above the large threshold on both scales, factor A (η² = .08, f = .33) sits between medium and large, and A × B (η² = .05, f = .26) is around medium by f and just under it by η². The two scales agree because f = .25 converts to η² of about .059 and f = .40 to about .138.
Treat these labels as a fallback, not a verdict. Lakens points out that Cohen developed the benchmarks for unrestricted populations, that they are inconsistent with ηp² in designs with covariates or repeated measures, and that using them for generalized eta squared is a last resort; he recommends comparing your effect to effects in the relevant literature instead. A .06 can be important in one field and negligible in another, which is a judgement about consequences and cost that no threshold makes for you.
Which ANOVA effect size should you report?
Lakens (2013) recommends reporting generalized eta squared, or both partial and generalized eta squared, rather than partial eta squared alone, because only generalized eta squared is comparable across designs. In a purely between-subjects design the two are the same number, so either label is accurate. Omega squared is a less biased estimate of the population effect; Lakens recommends it when it is easily computed, and notes the difference from η² is small and shrinks as N grows. The effectsize documentation adds that ω² and epsilon squared are the better choices in small samples, and that epsilon squared is analogous to adjusted R² and is less biased than ω².
Assumptions and common mistakes
- Not naming the version. A bare “η² = .17” is ambiguous when a design has several factors; write η² or ηp² explicitly.
- Applying the .01/.06/.14 labels to ηp² from a repeated-measures or covariate design. Lakens flags exactly this as inconsistent.
- Treating η² as a population value. It is the variance explained in the sample and is uncorrected; ω² or epsilon squared correct for this.
- Ignoring the sums-of-squares type. In unbalanced designs the sums of squares depend on the model’s type. The effectsize documentation says its results follow the model’s default type (for aov, type 1) and that type 3 needs effects coding or orthogonal contrasts.
- Comparing effect sizes across designs without checking they measure the same thing, for example partial values from a one-factor study and a four-factor study.
The usual ANOVA assumptions (independent observations, approximately normal residuals, similar variances across groups) apply to the effect sizes too, because they are computed from the same sums of squares; see assumptions of two-way ANOVA. When an interaction is present, a main-effect size can mislead, so read the interaction before the main effects, as in factorial analysis of variance.
How to report ANOVA effect size (APA-style example)
Report the F statistic with both degrees of freedom, the p-value and the effect size with its label. For factor B in the example:
There was a significant main effect of B, F(2, 54) = 5.63, p = .006, ηp² = .17.
For a result with an interval, report the confidence interval alongside it. The effectsize package computes intervals for these indices by the noncentrality-parameter (pivot) method. For wider reporting conventions see reporting statistics in APA style.
How to calculate effect size for ANOVA in Excel, R or Python
Excel
Put SSeffect in B2, SSerror in B3, SStotal in B4, dfeffect in B5, MSerror in B6 and then:
eta squared: =B2/B4
partial eta squared: =B2/(B2+B3)
omega squared: =B5*(B2/B5-B6)/(B4+B6)
Cohen's f: =SQRT((B2/(B2+B3))/(1-B2/(B2+B3)))
The sums of squares come from an ANOVA table, for example from the Data Analysis ToolPak.
R
library(effectsize)
m = aov(y ~ A * B, data = d)
eta_squared(m, partial = TRUE)
omega_squared(m)
epsilon_squared(m)
cohens_f(m)
These functions exist in the effectsize package; partial defaults to TRUE, and the generalized argument names the observed (non-manipulated) variables to compute generalized eta squared.
Python
import statsmodels.api as sm
from statsmodels.formula.api import ols
model = ols('y ~ C(A) * C(B)', data=df).fit()
aov = sm.stats.anova_lm(model, typ=2)
ss_err = aov.loc['Residual', 'sum_sq']
ms_err = ss_err / aov.loc['Residual', 'df']
ss_tot = aov['sum_sq'].sum()
aov = aov.drop('Residual')
aov['eta_sq'] = aov['sum_sq'] / ss_tot
aov['partial_eta_sq'] = aov['sum_sq'] / (aov['sum_sq'] + ss_err)
aov['omega_sq'] = aov['df'] * (aov['sum_sq'] / aov['df'] - ms_err) / (ss_tot + ms_err)
print(aov)
This applies the formulas above directly to the statsmodels table; see ANOVA in Python with statsmodels for fitting the model.
Frequently asked questions
What is a good effect size for ANOVA?
Cohen’s conventions label η² of about .01, .06 and .14 as small, medium and large, and f of .10, .25 and .40. They are a fallback; Lakens (2013) recommends comparing with effects in your own literature.
What is the difference between eta squared and partial eta squared?
Eta squared divides SSeffect by SStotal; partial eta squared divides it by SSeffect plus SSerror. They are equal in a one-way ANOVA and differ when a design has several factors.
How do you measure effect size for a two-factor ANOVA?
Compute the index separately for each main effect and for the interaction, using that effect’s sum of squares and the shared error term, as in the worked example above.
Is omega squared better than eta squared?
Omega squared is a less biased estimate of the population effect. The difference from eta squared is small and shrinks as the sample grows, so it matters most in small samples.
How do I convert partial eta squared to Cohen’s f?
Use f² = ηp² / (1 − ηp²) and take the square root. For ηp² = .172, f² = .208 and f = .456.
Sources
- 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. doi:10.3389/fpsyg.2013.00863 (read 2026-11-14).
- effectsize R package documentation, eta_squared reference page: easystats.github.io/effectsize (read 2026-11-14).
Data: ANOVA effect sizes for the worked 2 x 3 example
Effect-size indices computed from the ANOVA table of an illustrative balanced 2 x 3 between-subjects design (N = 60, SS error 432, df error 54) using the formulas in Lakens (2013) eqs 11-15; the example values are invented and every figure is computed from them.
| Effect | F | Eta squared | Partial eta squared | Omega squared | Partial omega squared | Cohen's f |
|---|---|---|---|---|---|---|
| A (df 1) | 6 | 0.08 | 0.1 | 0.066 | 0.077 | 0.333 |
| B (df 2) | 5.625 | 0.15 | 0.172 | 0.122 | 0.134 | 0.456 |
| A x B (df 2) | 1.875 | 0.05 | 0.065 | 0.023 | 0.028 | 0.264 |
Sources
- Lakens (2013) Calculating and reporting effect sizes, Frontiers in Psychology 4:863 (accessed 2026-11-14)
- effectsize R package: eta_squared reference (accessed 2026-11-14)
Compiled by CASRAI from the sources above. Free to reuse under CC BY 4.0 with attribution.







