Matematické Fórum

Ahoj,

jsem teď hodně zaneprázdněn, takže si dovolím postnout jednu odpověď na tvoji otázku, kterou jsem našel, aniž bych ji nějak analyzoval.

> Does anyone know if a module already exists to run the Scheirer-Ray-Hare
> extension to Kruskal-Wallis in R, in order to run a non-parametric
> equivalent to a two-way ANOVA?

I had a look at the JSTOR copy of the original Biometrics paper, and tend to say that I hope not... Amazing what you could get away with in the 70s!

At the very least, note that all of the test statistics are being evaluated under the _global_ null hypothesis, i.e. no A effect, no B effect, no interaction. Testing for significance of A in the presence of a strong effect of B is not what the test does, so it is not a replacement for two-way ANOVA. You need a stratified test like Friedman's or van Elteren's test for that.

That being said, the procedure it is just a decomposition of the K-W test for the effect of A*B, and you can easily compute the relevant decomposition of the SSD from an anova() of a linear model for the ranks.