One-Way ANOVA Calculator
Compute the F-statistic for a one-way ANOVA from the treatment and error sums of squares and their degrees of freedom. The standard test for comparing the means of three or more groups simultaneously — used in experimental design, A/B/C testing, and any multi-group comparison.
Last updated: May 2026
Compare with similar
About this calculator
One-way ANOVA tests whether the means of k groups (k ≥ 3) are all equal under the null hypothesis. The F-statistic is the ratio of two variance estimates: F = MS_treatment / MS_error = (SS_treatment/df_treatment) / (SS_error/df_error). Under H₀ both quantities estimate the same population variance, so F ≈ 1; if group means actually differ, the treatment mean square inflates and F grows large. SS_treatment (sum of squares between groups) measures the variability explained by group membership: SS_treat = Σ nⱼ (x̄ⱼ − x̄)², where x̄ⱼ is the j-th group mean, x̄ is the grand mean, and nⱼ is the size of group j. SS_error (within-group sum of squares) measures unexplained variability: SS_error = Σⱼ Σᵢ (xᵢⱼ − x̄ⱼ)². Degrees of freedom: df_treat = k − 1 (one less than number of groups), df_error = N − k (total sample size minus number of groups). Variables: SS_treatment is the between-groups sum of squares, SS_error is the within-groups (residual) sum of squares. Edge cases: the F-statistic must be non-negative; very small F values (close to 0) suggest groups are too similar (rare for the right test). Assumes (1) independence of observations, (2) normality within each group, (3) homogeneity of variance across groups (Levene’s or Bartlett’s test diagnoses violations). For non-normal data use Kruskal-Wallis instead. For heteroscedastic groups use Welch’s ANOVA or the Brown-Forsythe modification. ANOVA tests only whether ANY groups differ, not which specific groups — for that, follow up with pairwise post-hoc tests (Tukey’s HSD, Bonferroni) that control the family-wise error rate.
How to use
Example 1 — Testing four teaching methods. Four classrooms (k = 4), 10 students per class (N = 40, df_error = 36, df_treat = 3). After computing sums of squares: SS_treatment = 240, SS_error = 720. MS_treat = 240/3 = 80; MS_error = 720/36 = 20. F = 80/20 = 4.0. ✓ Compare to F(3, 36) critical value at α = 0.05 ≈ 2.87. Since 4.0 > 2.87, reject H₀ — at least one teaching method produces different mean outcomes. Follow up with Tukey’s HSD to identify which specific pairs differ. Example 2 — Three fertilizers, small experiment. Three plots (k = 3), 8 measurements per plot (N = 24, df_error = 21, df_treat = 2). SS_treatment = 18.5, SS_error = 84.0. MS_treat = 18.5/2 = 9.25; MS_error = 84.0/21 = 4.0. F = 9.25/4.0 ≈ 2.31. ✓ Compare to F(2, 21) critical value at α = 0.05 ≈ 3.47. Since 2.31 < 3.47, fail to reject H₀ — no statistically significant evidence that the fertilizers differ in their effects. The p-value is approximately 0.12. Either the effects are genuinely similar, or the experiment lacks power (only 8 plots per group); a larger sample might reveal differences if they exist.
Frequently asked questions
Why use ANOVA instead of multiple t-tests?
Running multiple pairwise t-tests on k groups inflates the family-wise Type I error rate dramatically. With k = 4 groups you have C(4,2) = 6 pairwise comparisons; at α = 0.05 each, the probability of at least one false positive is 1 − (1 − 0.05)⁶ ≈ 0.26, more than five times the nominal rate. ANOVA controls this by testing all groups simultaneously with a single test, keeping the overall Type I error at α. If ANOVA is significant, then pairwise post-hoc tests (Tukey’s HSD, Bonferroni-corrected t-tests, Scheffé) identify which specific groups differ while maintaining the family-wise error rate. ANOVA is more powerful than Bonferroni-corrected t-tests when the groups share a common variance, because it pools all observations to estimate that variance. For more complex designs (multiple factors, repeated measures, covariates), generalised linear models extend the basic one-way ANOVA framework.
What does a significant F-statistic actually tell you?
A significant F (p < α) tells you that at least one group mean differs from at least one other group mean — but not which ones or by how much. It is an omnibus test that detects the existence of differences without identifying them. After a significant F, run post-hoc tests to find the specific pairs: Tukey’s HSD is the most popular for all pairwise comparisons because it precisely controls the family-wise error rate; Bonferroni is simpler but more conservative; Scheffé works for any contrast including complex linear combinations of means. Without post-hoc analysis, a significant ANOVA tells you ‘there is something interesting here’ but stops short of saying what. Also, F-statistics measure existence of effects, not their magnitude — pair with eta-squared η² = SS_treatment / SS_total to get an effect size.
What are the assumptions of one-way ANOVA?
Three main assumptions: (1) independence of observations within and between groups; (2) normality of residuals within each group — for moderately large groups (n ≥ 25) the Central Limit Theorem makes this robust; (3) homogeneity of variances across groups (homoscedasticity) — tested with Levene’s or Bartlett’s test. Violations: heavy non-normality with small groups requires non-parametric Kruskal-Wallis; heteroscedasticity requires Welch’s ANOVA, which doesn’t assume equal variances; non-independent observations (paired, clustered, time-series) need mixed-effects or repeated-measures ANOVA. The classic ‘F-test is robust to moderate violations’ folklore holds best when group sizes are balanced (equal n per group); unbalanced designs are more sensitive to assumption violations. For very small groups (n < 10) the F-test is sensitive to all three assumptions and should be checked carefully or replaced with Kruskal-Wallis.
What are the most common mistakes people make with ANOVA?
The first is using ANOVA when the data has only two groups; that is just a t-test (and the F-statistic is exactly t²). The second is interpreting a significant F as evidence about which specific groups differ; F is omnibus, so follow up with post-hoc tests. The third is running ANOVA on dependent observations (e.g., the same subjects measured under different treatments); use repeated-measures ANOVA instead. The fourth is ignoring violations of homogeneity of variance; severely heteroscedastic groups need Welch’s ANOVA. The fifth is reporting only F and p without effect size (η², ω²) and group means; ANOVA hides the magnitude and direction of differences. The sixth is doing multiple ANOVAs across many outcomes without correcting for multiple testing; this inflates false positives just like multiple t-tests would. The seventh is treating non-significant F as proof of no difference; it could just mean the experiment lacks power.
When should I not use this calculator?
Skip it for two-group comparisons; use a t-test. Avoid it when the assumption of normality is severely violated with small groups; use the non-parametric Kruskal-Wallis test instead. It is the wrong tool for paired/repeated-measures data; use repeated-measures ANOVA or mixed-effects models. Do not use it for designs with multiple categorical factors (two-way, three-way ANOVA), which require expanded sum-of-squares decompositions for main effects and interactions; this calculator handles only the one-way case. Skip it for nested or hierarchical data (students within classrooms, patients within hospitals) where you need multilevel modelling. And do not use one-way ANOVA when group sizes are very unequal (e.g., one group of 5 vs another of 500) without careful diagnostics — heteroscedasticity becomes much more harmful in unbalanced designs.