Interactive statistics

56 free interactive lessons: from mean and median to A/B tests and Bayes. Drag the charts and build intuition. No sign-up.

Module 11: Analysis of variance

After ANOVA: which groups actually differ

ANOVA only says "there is a difference somewhere". But the business needs "WHICH group is better". The naive answer — run t-tests for all pairs — leads straight into the multiple comparisons trap. That is why special post-hoc procedures follow ANOVA.

difference = 0A−Bnot sig.A−CsignificantB−Cnot sig.

Each row is a pair’s difference of means with an interval. Missed zero (green) → significant. Switch the Tukey correction off — the intervals narrow and more pairs turn "significant": that is the inflation of false findings under uncorrected pairwise tests.

ANOVA said "there is a difference somewhere". But between WHICH of the three groups? The chart shows confidence intervals of the pairwise differences (A−B, A−C, B−C). The rule is simple: if an interval does not cover zero (green) — the difference is significant; if it covers it (gray) — it is not. Move the between-group and within-group spreads — the intervals respond.

What it means

The proper multi-group analysis: ANOVA → if significant → a corrected post-hoc (Tukey). That way you neither drown in pairwise tests nor miss which group truly stands out.

The same in A/B/C/D tests: first check whether the variants differ at all, then corrected comparisons against the control. Otherwise, with many variants the "winner" is often just chance.

Where it shows up

Agronomy and medicine are the birthplace of post-hocs: compare several fertilizers or doses and figure out which one is actually better without harvesting false differences.

Product work: comparing several landing page versions or pricing grids — ANOVA/regression at the gate, then corrected pairwise comparisons to pick the variant.

Definitions
Post-hoc
comparisons after a significant ANOVA to find out which groups actually differ.
Tukey's test (HSD)
a post-hoc procedure for all pairs that holds the overall false-positive risk at the chosen level.
Family-wise error (FWER)
the probability of at least one false finding among all comparisons; this is what the corrections control.
Two-step logic
first ANOVA (is there a difference), then the post-hoc (where exactly) — preserves power.
When the method lies (assumptions)

Post-hocs inherit ANOVA's assumptions (roughly normal residuals, similar group spread, independence). When violated — nonparametric analogues (Kruskal–Wallis + Dunn).

Controlling the family-wise error reduces the power of each individual comparison: the honest price for the right to test many pairs at once.

Next →
Enjoying the materials?

This site is built by the DataSlice Telegram channel: data analytics in plain words — real cases, metrics, careers. The channel is in Russian.

Open the Telegram channel