ANOVA (Analysis of Variance) tests whether three or more population means are equal by splitting the total variation into variation between groups and within groups and comparing them with an F-test; its significance is that it compares many groups at once without inflating the error rate.
Basic idea
\[SST = SSB + SSW,\qquad F = \frac{SSB/(k-1)}{SSW/(N-k)}\]
If group means differ only by chance, between-group and within-group variances are similar and \(F \approx 1\). A large F (greater than the table value) means the group means really differ, so \(H_0: \mu_1 = \dots = \mu_k\) is rejected.
ANOVA table: Source (Between, Within, Total) | SS | df (\(k-1\), \(N-k\), \(N-1\)) | MS = SS/df | F = MSB/MSW.
Types: one-way (one factor), two-way (two factors with or without interaction), and repeated-measures ANOVA.
Significance
- Controls Type I error: doing many t-tests (e.g. 10 tests for 5 groups) raises the chance of at least one false positive well above 5%; one ANOVA keeps it at \(\alpha\).
- Handles several groups and factors: two-way ANOVA also tests interaction (e.g. whether a fertiliser works differently on different soils).
- Basis of experimental design: used in completely randomised, randomised block and Latin square designs.
- Wide use: agriculture (crop yields), medicine (drug doses), industry (machines or shifts), education (teaching methods), marketing (ad versions), and ML (ANOVA F-score for feature selection).
- Leads to further analysis: if significant, post-hoc tests (Tukey, Bonferroni) find which groups differ.
Assumptions: independence, normality within groups, and homogeneity of variances.