One-Way ANOVA Calculator
One-way ANOVA: compare means across 2 or more groups. Outputs F-statistic, degrees of freedom, p-value, eta-squared effect size + per-group descriptive statistics.
One-Way ANOVA Calculator
How to use the ANOVA Calculator
Enter your groups
One group per line; values separated by commas or spaces. Minimum 2 groups (with 2+ values each), but ANOVA is most useful for 3+ groups (use t-test for exactly 2). Groups can have different sample sizes; the math accounts for this.
Read the F-statistic + p-value
F is the ratio of "between-group variability" to "within-group variability." Large F → likely real differences between groups. p-value tells you the probability of seeing this F (or larger) if all group means were truly equal. p < 0.05 typically means at least one group differs.
Check η² for effect size
Eta-squared (η²) is the proportion of total variance explained by group membership. Cohen's (1988) benchmarks: 0.01 small, 0.06 medium, 0.14 large. Important: with large samples F can be highly significant but η² small — a real-but-tiny effect. Always report both p and η².
Run post-hoc tests if significant
ANOVA significance tells you "at least one group differs" but NOT which one(s). To find the differing pair(s), run post-hoc tests: Tukey HSD (most common, all pairwise comparisons); Bonferroni (conservative); Dunnett (each group vs control). Most stats software automates these.
ANOVA — comparing means across 3 or more groups without inflating false positives
Analysis of Variance (ANOVA), developed by Ronald Fisher in the 1920s, solves a specific problem: when you have 3 or more groups to compare, running pairwise t-tests inflates the false-positive rate. Comparing 4 groups via 6 separate t-tests at α = 0.05 each gives a family-wise false-positive rate of about 26% (1 − 0.95⁶, treating the six tests as independent), not 5%. ANOVA provides a SINGLE omnibus test that maintains the overall α level. If ANOVA is significant, you then run post-hoc tests (Tukey HSD, Bonferroni) with proper multiple-comparison correction to find which specific groups differ. The math decomposes total variance into "between-group" and "within-group" components; the F-statistic is the ratio of these. Large F → group differences are larger than within-group noise → likely real differences.
Why ANOVA + not just multiple t-tests
With k groups, there are k(k-1)/2 possible pairwise comparisons. For k=4 groups: 6 comparisons. At α = 0.05 per comparison, the family-wise error rate (probability of at least one false positive) is 1 − (1-0.05)^6 ≈ 26% if the comparisons were independent. With k=10 groups (45 comparisons): family-wise error ≈ 90%. ANOVA controls this by running ONE test that asks "is there any difference among ALL groups?" — if no, stop. If yes, follow up with controlled post-hoc tests. This is why an omnibus test comes before pairwise comparisons when 3+ conditions are compared — the pattern Fisher set out in Statistical Methods for Research Workers (1925) and that the NIST/SEMATECH e-Handbook (§7.4.3) still lays out step by step.
Comparing 4 groups via 6 separate t-tests inflates the false-positive rate from 5% to about 26%, if the tests are treated as independent. ANOVA's single omnibus test controls this. Run ANOVA first, then post-hoc tests.
What the F-statistic actually measures
F = MS_between / MS_within. MS_between ("mean square between groups") measures how much group means differ from the overall mean — high when groups differ substantially. MS_within ("mean square within groups") measures how much individual values differ from their own group mean — measures noise inherent to the data. F ratio interpretation: F=1 means between-group variability ≈ within-group noise (no signal). F=2 means signal is twice the noise. F=10+ means strong signal. The p-value translates F into "probability of seeing this F if all groups were truly identical."
The ASEAN multivariate testing context
Multi-armed product experiments (A/B/C/D testing) are where a team comparing more than two variants meets this test. For continuous metrics (revenue per user, session length, time-to-purchase): one-way ANOVA is the right omnibus test. For categorical metrics (conversion rate): chi-square test of independence is the analog. Common ASEAN cross-country comparison: testing whether some metric (engagement, ARPU, retention) differs across SG/MY/PH/ID/TH/VN markets. ANOVA tells you "is there any country-level difference?" Tukey HSD post-hoc tells you specifically which countries differ. Watch out: with the sample sizes a consumer platform accumulates, even trivial differences become statistically significant. Always check η² to confirm differences are practically meaningful, not just statistically detectable.
Rothamsted fertiliser plots, F = t² for two groups, and the 26% that six t-tests cost you
Ronald Fisher developed ANOVA for agricultural field trials at Rothamsted Experimental Station and set it out for general use in Statistical Methods for Research Workers (1925).
F = MS_between / MS_within. Ratio of "signal" (group differences) to "noise" (within-group variability).
ANOVA replaces multiple t-tests when comparing 3+ groups — six independent t-tests at α = 0.05 would carry a family-wise false-positive rate near 26%.
η² (eta-squared): proportion of variance explained. Cohen's (1988) benchmarks — f = 0.10, 0.25, 0.40 — correspond to η² ≈ 0.01, 0.06 and 0.14.
Significant ANOVA → run post-hoc tests (Tukey HSD, Bonferroni, Dunnett) to find which groups differ.
Assumes normality + equal variances. For unequal variances: Welch's ANOVA. For non-normal: Kruskal-Wallis.
F-distribution is right-skewed, always positive. Large F always indicates more between-group variation than within-group.
For 2 groups: ANOVA is mathematically equivalent to a t-test. F = t². Use t-test directly for clarity.
"One-way" = one categorical predictor. "Two-way" = two predictors with interaction. "Three-way" + factorial designs handle complex experiments.
ASEAN A/B/C/D testing: ANOVA + Tukey HSD reveals which specific variants drive significant differences, not just "any difference exists."
Frequently Asked Questions
-
Family-wise error rate balloons. With k groups, k(k-1)/2 pairwise tests; at α=0.05 each, the probability of AT LEAST ONE false positive, treating the comparisons as independent, is 1 − (0.95)^n_comparisons. 4 groups (6 comparisons) → about 26%. 10 groups (45 comparisons) → about 90%. Pairwise tests that share groups are not truly independent, so the exact figure differs, but the rate still climbs well above 5%. ANOVA gives ONE test maintaining overall α = 0.05. Then post-hoc tests apply proper multiple-comparison correction (Tukey HSD, Bonferroni) on the surviving question.
-
One-way ANOVA (this tool): one categorical predictor (e.g., "Diet Type" with 4 levels). Two-way ANOVA: two predictors + interaction (e.g., "Diet Type" × "Exercise Level"). Two-way tests both main effects AND whether the effect of one predictor depends on the level of the other (interaction effect). For interaction-effect testing, use specialised software (R aov(), Python statsmodels, SPSS).
-
Standard ANOVA assumes equal variances (homoscedasticity). Welch's ANOVA doesn't — uses adjusted degrees of freedom to handle unequal variances. Robust to violations. Detection: Levene's test or Bartlett's test on group variances. Practical rule: if the largest standard deviation is more than 2× the smallest, use Welch's ANOVA instead. R aov() does standard; oneway.test() with var.equal=FALSE does Welch's. This tool currently does standard ANOVA; for unequal variances, validate with R or Python.
-
Tukey HSD: most common; balanced + powerful; all pairwise comparisons; controls family-wise error. Bonferroni: very conservative; α/n_comparisons; easy to apply manually. Scheffé: flexible for arbitrary contrasts beyond pairs. Dunnett: best when comparing each treatment to a control (don't need all pairs). Default choice: Tukey HSD unless you have a specific reason for another. This calculator doesn't run post-hoc tests directly — use R (TukeyHSD function), Python (statsmodels), or SPSS for them.
-
Kruskal-Wallis = non-parametric alternative to ANOVA. Use when: data is severely skewed, has heavy outliers, or is on an ordinal scale (rankings, Likert scales). Tests whether group medians differ rather than means. Less powerful than ANOVA when ANOVA's assumptions hold, but more robust when they're violated. Available in R (kruskal.test), Python (scipy.stats.kruskal), SPSS, JASP.
-
Standard APA format: "F(df_between, df_within) = X.XX, p = .XXX, η² = X.XX". Example: "Diet had a significant effect on weight loss, F(3, 56) = 5.42, p = .002, η² = .23." Always include: (1) F-statistic + both degrees of freedom; (2) exact p-value (or "< .001"); (3) effect size (η², ω², or partial η²); (4) descriptive stats (means + SDs) for each group; (5) post-hoc comparison results if applicable.
-
Yes — for exactly 2 groups, F equals the square of the pooled-variance (Student) t-statistic, and the p-values agree. If your t-test result does not square to this F, it is usually because: (1) the t-test was Welch's, which does not pool variances, while ANOVA does; (2) rounding; (3) the two calculations were fed different data. For 2-group comparisons, use the t-test directly — it's mathematically equivalent and reporting "F" instead of "t" is unconventional.
-
Cohen's conventions (Statistical Power Analysis for the Behavioral Sciences, 2nd ed., 1988) are stated for the effect size f — 0.10 small, 0.25 medium, 0.40 large — which convert to η² = f²/(1 + f²) ≈ 0.01, 0.06 and 0.14. Interpretation: η² = 0.14 means group membership explains 14% of total variance. Cohen himself called these benchmarks conventions for when a field has no better yardstick; what counts as a meaningful difference is a judgement about your subject, not a property of the statistic.
-
No. All calculations run in your browser via JavaScript. Open DevTools → Network and confirm zero outbound requests. Data stays on your device. Safe for confidential research + business data.
-
Pair with: the T-Test Calculator for 2-group comparisons; the Chi-Square Calculator for categorical data; the Sample Size Calculator for pre-test power analysis; the Linear Regression Calculator for continuous predictors. External: R (aov, TukeyHSD), Python (scipy.stats.f_oneway, statsmodels), SPSS, JASP, JMP — all support full ANOVA + post-hoc workflows.
Method & sources
How it computes
Fixed-effects one-way analysis of variance: SS_between = Σ nᵢ(x̄ᵢ − x̄)², SS_within = Σ Σ (xᵢⱼ − x̄ᵢ)², F = [SS_between/(k − 1)] / [SS_within/(N − k)], with the p-value P(F_{k−1, N−k} ≥ F) taken from the F distribution via the regularised incomplete beta function I_x(d₁/2, d₂/2), x = d₁F/(d₁F + d₂). Effect size is η² = SS_between/SS_total; per-group n, mean, sample SD (n − 1), min and max are reported alongside.
What this tool implements
- Classical (Fisher) fixed-effects one-way ANOVA with a pooled within-group variance — assumes independent observations, approximately normal groups and equal variances; Welch's ANOVA, Kruskal–Wallis and post-hoc tests (Tukey HSD, Bonferroni, Dunnett) are described on the page but NOT computed.
- Unequal group sizes are handled by weighting each group's squared mean deviation by its nᵢ, as in the NIST/SEMATECH hand-calculation (§7.4.3.4); k ≥ 2 groups with nᵢ ≥ 2 each.
- p-value from the F distribution computed with the Numerical Recipes gammln / betacf continued-fraction algorithm for the incomplete beta function (tolerance 3 × 10⁻⁷); p below 10⁻⁴ is displayed as '< 0.0001'.
- η² interpreted against Cohen's (1988) conventions f = 0.10 / 0.25 / 0.40, i.e. η² ≈ 0.01 small, 0.06 medium, 0.14 large.
Sources
- Fisher RA. Statistical Methods for Research Workers. Edinburgh: Oliver and Boyd; 1925. Chapters VII–VIII: intraclass correlation and the analysis of variance.
- NIST/SEMATECH e-Handbook of Statistical Methods. Section 7.4.3: Are the means equal? (one-way ANOVA, the ANOVA table and the F-test; §7.4.3.4 hand calculations). https://www.itl.nist.gov/div898/handbook/prc/section4/prc43.htm
- Press WH, Teukolsky SA, Vetterling WT, Flannery BP. Numerical Recipes: The Art of Scientific Computing. 3rd ed. Cambridge: Cambridge University Press; 2007. §6.1 Gamma function, §6.4 Incomplete beta function and the F-distribution.
- Cohen J. Statistical Power Analysis for the Behavioral Sciences. 2nd ed. Hillsdale, NJ: Lawrence Erlbaum Associates; 1988. Chapter 8: the F test on means, effect size f = 0.10, 0.25, 0.40.
What can make this go out of date
- None at runtime — the F distribution is computed in the browser; no tables or published series are consulted.
Pick up where you left off
Stored only in this browser — never sent to our servers.