ANOVA Calculator

Ronald A. Fisher introduced analysis of variance in a 1918 paper on the correlation between relatives on the supposition of Mendelian inheritance, then developed it into a general statistical method in Statistical Methods for Research Workers (1925) and The Design of Experiments (1935). The central idea is to partition the total variance of the data into a between-groups component and a within-groups component, then compute their ratio, the F-statistic, named after Fisher by George Snedecor (1934). Under the null hypothesis that all group means are equal, F follows an F-distribution with the appropriate degrees of freedom, and a large F leads to rejection. A single omnibus test, run once, controls the family-wise error rate at α, which is why ANOVA is preferred over running multiple t-tests when comparing three or more groups.

The choice of design follows the structure of your data. One-way ANOVA compares three or more independent groups on a single categorical factor: three treatment arms in a randomised trial, four educational interventions, five varieties of wheat in an agronomy trial. Two-way (factorial) ANOVA crosses two factors (for example treatment × dose, sex × condition, or material × temperature) and decomposes the variance into a main effect of A, a main effect of B, and the A × B interaction. A significant interaction means the effect of one factor depends on the level of the other, and is usually the most clinically informative term. The calculator's two-way mode assumes a balanced design (equal cell sizes); for unbalanced data, switch to R's car::Anova() with type-II or type-III contrasts.

Repeated-measures ANOVA is the within-subjects analogue: each subject is measured under every condition or at every time point, and between-subject variability is partitioned out as a separate variance component. This dramatically increases statistical power compared to a between-subjects design (fewer subjects are needed to detect the same effect), but it introduces the sphericity assumption: the variances of the pairwise differences between conditions must be equal. Mauchly's test (1940) checks sphericity; when it fails, the Greenhouse-Geisser (1959) and Huynh-Feldt (1976) corrections shrink the degrees of freedom by an estimated ε ∈ [1/(k − 1), 1] to restore valid inference. The calculator computes both ε values from the contrast-transformed covariance matrix and reports the corrected p-values automatically.

Three assumptions sit behind every ANOVA: independence of observations, normality of residuals within each group, and homogeneity of variances across groups (or sphericity for repeated measures). ANOVA is robust to mild violations, especially with equal cell sizes, but when variances differ substantially Welch's ANOVA (Welch 1951) is the modern default; it does not pool the within-group variance and adjusts the degrees of freedom using the Welch-Satterthwaite formula. When normality fails, Kruskal-Wallis (1952) is the non-parametric counterpart for one-way designs and Friedman's test (1937) for repeated measures. The calculator flags violations in its assumptions panel and suggests the appropriate remedy.

A significant omnibus F tells you that at least one group mean differs from the others, but not which. Post-hoc pairwise comparisons answer that question while controlling the family-wise error rate. Tukey's HSD (Tukey 1949) is the standard for balanced designs because it controls the family-wise error rate exactly under the studentized range distribution. Bonferroni (Dunn 1961) multiplies each raw p-value by the number of comparisons and is the most conservative valid procedure. Holm-Bonferroni (Holm 1979) is a step-down variant that is less conservative without sacrificing rigor. Sidak (1967) assumes independence and is slightly less conservative than Bonferroni. Dunnett's test (1955) is the right choice when every group is compared only to a single control. The calculator returns Bonferroni, Holm, and Sidak adjusted p-values in a sortable table so you can pick by audience.

ANOVA feeds directly into downstream tools. To run the prerequisite Levene or Brown-Forsythe test of variance equality from raw data or summary stats, use the variance calculator with its F-test for equal variances. For a two-group comparison rather than three or more, use the two-sample t-test calculator. To convert the F-statistic into a Cohen's d or correlation r effect size for meta-analysis, use the effect size calculator. To plan sample size for an ANOVA design given a target Cohen's f and power, use the sample size calculator. For full data extraction, assumption checking, and APA-formatted manuscripts from a PhD statistician, our statistical analysis service covers everything end-to-end.

Reporting ANOVA results in APA format requires four elements together: the F statistic with both degrees of freedom, the exact p value, an effect size (eta squared or the less biased omega squared), and the follow-up pairwise comparisons with the correction method named. A significant omnibus F alone only says that some difference exists somewhere; reviewers expect you to say which groups differ. Using three example groups (n = 8 each), a complete write-up looks like this:

A one-way analysis of variance was conducted to compare the outcome across the three groups (N = 24). The omnibus test was statistically significant, F(2, 21) = 54.44, p < .001, η² = .838 (ω² = .817), a large effect. Bonferroni-corrected pairwise comparisons showed that all three groups differed from one another (all p < .001), with means of 14.62, 23.50, and 31.50 respectively.

The ANOVA table in APA format has horizontal rules only and lists Source, SS, df, MS, F, and p, with Between groups, Within groups, and Total rows. The APA Report button above downloads exactly that: a Word document with the Results paragraphs, the APA-ruled ANOVA summary table, group means in the table note, a plain-English interpretation of the F test and effect sizes, and a reporting checklist covering the mistakes committees flag most, such as reporting F without both degrees of freedom or claiming two specific groups differ without a corrected pairwise test.