In another network, senior author Adam Claridge-Chang publicly added this insightful and pungent commentary:
"Here's a dirty secret about ANOVA: it tests a null hypothesis that nobody cares about. When you run a one-way ANOVA, you're testing whether "all group means are equal." But even if you reject this hypothesis, you learn nothing about which groups differ, in which direction, or by how much. So you embark on a second analytical step: multiple two-group comparisons. A modest six-group experiment suddenly requires testing 15 hypotheses. To manage this multiplicity, you apply corrections like Bonferroni, which undermine your statistical power. What you posed as a focused research question has sprawled into a complex web of subsidiary tests, forced by the ANOVA ritual."
"Our new preprint, "Getting over ANOVA: Estimation graphics for multi-group comparisons," makes the case for a better approach. Estimation statistics encourages you to compare each test group to a single control, focusing on the effect sizes that actually matter. A six-group experiment focuses attention on just five effect sizes with confidence intervals, showing magnitude and precision directly."
"The preprint introduces estimation methods for a range of multi-group designs: repeated-measures experiments, 2×2 factorial designs, binary outcome data, and mini-meta analysis for internal replicates. Each can replace data-analysis practices used in thousands of studies every year."
#ANOVA #statistics